MTS 363 - OPERATIONS RESEARCH ... Dr.I.A.

Osinuga
Recommended Texts
1.

Operations Research (an introduction) Handy A. Taha (3rd Edition)

2.

Optimization (Theory and applications) S.S. Rao

3.

Principles of Mathematical Modelling Ciive, L. Dym and Elizabeth, Slvey

4.

Introduction to Management Science (2nd Edition) Bernard W. Taylor

5.

An introduction to Management Science (7th Edition) D.R. Anderson et al

6.

Quantitative Techniques T. Lucey

7.

Apply Mathematics for Business, Economics and the Social Sciences (4 th Edition)
F.s. Budnik

Topics Covered in the Lecture notes:

The Nature of OR; Origin of OR

Network Analysis & Project Control

Models and Modelling in OR

Inventory Problem

Linear Programming Models

Game Theory

Formulation

Dynamic Programming

Solving an LPM

Integer Programming

Graphical method

Decision Theory

Simplex Method

Duality Theory and Sensitivity Analysis

Application of LPM
-

Transportation Problem

Assignment Problem

THE NATURE OF OPERATIONS RESEARCH

Definition
The phrase operation research is used to describe the application of scientific
method (observe, model, predict) to the operation and management of organizations
involving people and other resources (for example, money and machinery).
The nature and origin of OR
The activity of OR really got under way in England and then in the United States,
during the second world war, when the motivation was the optimization of the allocation
of scarce military resources. The application of science to warfare dates back to ancient
times and it involves such luminous names as Archimedes and Leonardo da Vinci. These
applications of science were devoted to developing ingenious new tools of warfare, new
armaments. The modern development of operations research was organized on the
principle of improving the utilization of existing stocks of resources (armament) by
careful application of the scientific methods. Lanchesters law represents an early effort
in the direction of present day operations research being the first modern attempt to
model the interaction of opposing armies at war.
The scope of OR has expanded in the last 30 years and it has been found in civil as
well as military settings. Some of the civil applications are to hospital management,
criminal justice system operation, and a variety of commercial enterprises. A hallmark of
OR is that it is directed toward achieving optimal solutions to problems. E.g., toward
finding the mix of products for a manufacturer that will maximize profits, towards
helping the manufacturer choose the right distribution of products among various outlets
locations so as to minimize transportation costs and so on.
Phases of OR study
The major phases through which an OR team will proceed in order to effect an OR
study include:
1.

Definition of the problem

2.

Construction of the model

3.

Solution of the model

4.

Validation of the model

5.

Implementation of the final results

Models and Modelling in OR

Definition: A model is a general term denoting any abstract or idealized representation of
a real life system or situation or simply idealized representation of a real life system or
situation. Models provide a concise framework for analyzing a decision problem in a
systematic manner and in this respect two basic components are essential for constructing
a model: the ***objective of the system and the constraints imposed on the system.
Types of OR Models
The most important of OR models is the symbolic or mathematical model. It
assumes that all the relevant variables, parameters and constraints as well as the objective
are quantifiable. Thus if x j, j = 1, 2, , n are the n decision variable of the problem under
study, and if the system is subject to m constraints, the general mathematical model can
be written in the form.
Optimize z = f (x1, xn)

(objective)

s.t.
g1 (x1, xn) b1
x1, x2, , xn 0.

i = 1, 2, m
constraints

In addition to mathematical model, simulation and heuristic models are used. Simulation
models initiate the behavior of the system over a period of time. This is achieved by
specifying a number of events that are points in time whose occurrence signifies that
important information pertaining to the behaviour of the system can be gathered.

LINEAR PROGRAMMING MODELS

Introduction
A linear programming is a mathematical technique applicable when the
relationship among variables can be expressed as directly proportional (Linear) functions.
Every linear programming problem consists of a mathematical statement called an
objective function. This function is to be either maximized or minimize, depending on the
nature of the problem. For example, profit would be maximized but cost would be
minimized. Also involved is a set of constrains equations which can either be in form of
inequalities or equalities of the difference resources available and the proportion of each
resource necessary to make a unit of the item of interest such as manufactured parts,
personal policies, inventories etc.
FORMULATION OF LINEAR PROGRAMMING MODELS
The usefulness of linear programming as tool for optimal decision making and
resource allocation is based on its applicability to many diversified decision problems as
determining the most profitable product mix, scheduling inventory, planning manpower
management etc. it has been used for pollution control, personal allocation, capital
budgeting and financial personnel selection.
The effective use and application require, as a first step the formulation of the
model when the problem is presented. The three basic steps in formulating a linear
programming are as follows:
Step 1
Identify the decision variables to be determined and express them in terms of algebraic
equations.
Step 2
Identify all the limitations or constraints in the given problem and then express them as
linear inequalities or equalities, in terms of above identified decision variables
Step 3
Identify the objective (criterion) which is to be optimized (maximized or minimized) and
express it as a linear function of the above defined decision variables.

We present below some illustrations on the formulation of linear programming

models in various situations drawn from different area of management.
Example 1: Production Planning Problem
A tailor has the following materials available. 16 square meters of cotton, 11
square meters of silk, and 15 square meters of wool. He can make out two products from
these three materials, namely dress and suite. A dress requires the following: 2 square
meters of cotton, 1 square meter of silk and 1 square meter of wool. A suite requires 1
square meter of cotton, 2 square meter of silk and 3 square meter of wool. If the gross
profit realized from a dress and as suite is respective N30 and N50. Formulate the above
as a linear programming model.
Solution:
The information needed to formulate the above problem is summarized in the table below
Table 1
Product

Dress

Suite

Material

Cotton

16

Silk

11

Wool

15

Profit

N30

N50

Material

Let x1 be number of dresses to be made

x2 be the number of suites to be made
Then the total profit z = 30x1 + 50x2 which is the objective function of the problem.
The following are constraints or limitations of the problem:
(a)

Only 16 sq. meter available for cotton to be used hence we have 2x 1 + x2 16

(b)

Limitation on silk would imply that x1 + 2x2 11

(c)

That on wool also means x1 + 3x2 15

(d)

And finally, at worst the tailor would make no garment implies that x1, x2 0

Rewriting all together we have their linear programming models to be as follows:

Maximize

z = 30x1 + 50x2

(objective function)

s.t.
2x1 + x2 16

(cotton constraint)

x1 + 2x2 111

(silk constraint)

x1 + 3x2 15

(wool constraints)

x1 , x 2 0

(non negative constraints)

Example 2: Cost Minimization Problem:

A manufacturer is to market a new fertilizer which is to be mixture of two
ingredients A and B. The properties of the two ingredients are as follows: ingredient A
contains 20% bone meal, 30% nitrogen, 40% lime and 10% phosphate and it cost N2.40
per kilogram. Ingredient B contains 40% bone meal, 10% nitrogen, 45% lime and 5%
phosphate, and it costs N1.60 per kilogram. Furthermore it is decided that:
(i)

The fertilizer will be sold in bags containing a minimum of 50 kilograms

(ii)

It must contain at least 12% nitrogen

(iii)

It must contain at least 6% phosphate

(iv)

It must contain at least 20% bone meal

Formulate the above problem as a LPM.

Solution:
Let x1 = number of kilogram of ingredient A
x2 = number of kilogram of ingredient B
The objective function is
Minimize z = 2.4x1 + 1.6x2
The following are constraints or limitations
(a)

Total weight constraints: x1 + x2 50

(b)

Bone meal constraints: 0.2x1 + 0.4x2 0.20

(c)

Nitrogen constraints: 0.3x1 + 0.1x2 0.12

(d)

Phosphate constraints: 0.1x1 + 0.05x2 0.06

(e)

Non negative constraints: x1, x2 0.

GRAPHICAL SOLUTION METHOD

One way of solving a linear programming problem after its formulation is the use
of graph. This is possible if the decision variables are not more than three. It is very
handy when the decision variables are two since a graph in 2 dimensions is easier to draw
than that in 3 dimensions; anything above three dimensions may be difficult.
Example 3:
Suppose a manufacturer produce 2 liquids, X and Y because of past sales
experience the market research estimates that at least as much Y as X is needed. The
manufacturing capacity of the plant allows for a total of 9 units to be manufactured. If
each unit of liquid X results in a profit of N2 and the profit for each unit of Y is N1, how
much of each should be produced to maximize profit?
Solution:
Let x be the amount of x to be produced
y be the amount of Y to be produced
We may state the given problem mathematically as follows.
Max z = 2x + y
s.t
2x y 0
x+y9
x, y 0
in the next section we develop a technique (the simplex algorithm) for solving such
problems in general, but we can solve this particular one geometrically. First we sketch
the set of point in IR2 that satisfy the set of constraints (shaded part of fig. 1). This region
is called the feasible region.

10
9
8
7
6
Series1

Series2

4
3
2
1
0
0

Figure 1
If a point is in the region, it satisfies all of the constraints and is called feasible, if it is not
in the region; it violates at least one of them and is called infeasible.
Different points in the infeasible region give different values of the objective
function. For example, the region (0, 0) is feasible and gives a profit 2 (0) 0 = 0.
Similarly, (1, 5) is feasible and gives profit of 2 (1) + 5 = 7, which is better than 0, we
seek the point or points of the feasible region that yield the maximal profit.
Corresponding to a fixed value of z, the set of solutions to the equation z = 2x + y
is the line (called the objective line) with the slope -2 and y intercept z. In other words, all
points along the line y = -2x + z correspond to the same z value. Remember that we are
interested only in those points that lie in the feasible region, and we want z to be as large
as possible (fig. 2)

15
14
13
12
11
10
9

Series1

8
7

Series2

6
5

Series4

Series3

4
3
2
1
0
0

Figure 2

2.0

THE SIMPLEX METHOD

Simplex method is a general purpose approach employed in solving linear

programming problems that are too large to be solved graphically. At this time, the
simplex method is by far the most widely used algebraic procedure for solving LPP.
Computer programs based on this method can routinely solve LPP with thousands of
variables and constraints.
2.1

An Algebraic Overview of the Simplex Method

We introduce the problem we will use to demonstrate the simplex method.

Lowtech industries import electronic components that are used to assemble two different
models of personal computers. One model is called the LT Deskpro computer and the
other model is called LT portable computer. Lowtech management is currently interest in
developing a weekly production schedule for both products.

The Deskpro generates a profit contribution of N50 per unit, and the portable
generates a profit contribution of N40 per unit. For next weeks production, a maximum
of 150 hours of assembly time can be made available. Each unit of the Deskpro requires 3
hours of assembly time, and each unit of the portable requires 5 hours of assembly time.
In addition, Lowtech currently has only 20 portable display components in inventory;
thus no more than 20 units of the portable may be assembled. Finally, only 300 square
feet of warehouse space can be made available for new production. Assembly of each
Deskpro requires 8 square feet of warehouse space; similarly, each portable requires 5
square feet.
Let x1 = no of units of the Deskpro assembled
x2 = no of units of the Portable assembled
The complete mathematical model for this problem is presented below:
Max 50x1 + 40x2
s.t
3x1 + 5x2 150

Assembly Time

1x2 20

Portable display

8x1 + 5x2 300

Warehouse capacity

Adding a slack variable to each of the constraints permits us to write the problem in
standard form.
Max 50x1 + 40x2 + 0s1 + 0s2 + 0s3
s.t.
3x1 + 5x2 + s1

150

1x2

20

8x1 + 5x2 + s3

300

x1, x2, s1, s2, s3 0

2.2

Basic Solution
To determine a basic solution, set n-m of the variables equal to zero, and solve the

m linear constraints equations for the remaining m variables. We shall refer to the n-m
variables set equal to zero as the non basic variables and the remaining m variables
allowed (to be non zero) as the basic variables.
A basic solution can either be feasible or infeasible. A basic infeasible solution is a
basic solution that also satisfies the non negativity condition otherwise it is infeasible.
For example, if we set x2 = 0 and s1 = 0 we obtain the following solution to the
three-equation, five variable set of linear equation determined by the Lowtech
constraints.
x1 = 50

s1 = 0

s3 = -100

x2 = 0

s2 = 20

The above solution is referred to as a basic solution for the Lowtech LPP but the
basic solution is not a feasible solution because s3 = -100. However, suppose we had
chosen to make x1 and x2 non basic variables by setting x1 = 0 and x2 = 0. Then the
complete solution corresponding to x1 = 0 and x2 = 0 is:
x1 = 0, x2 = 0, s1 = 150, s2 = 20, s3 = 300.
This solution is a basic solution since it was obtained by setting two of the
variables equal to zero and solving for the other three variables. Moreover, it is a basic
feasible solution since all of the variables are greater than or equal to zero.
2.3

Setting the Initial Simplex Tableau

To provide a convenient means for performing the calculation required by the

simplex solution procedure, we will first develop what is referred to as the initial simplex
tableau.
We adopt the general notation below in the tableau for representation of a linear
program.
cj = objective function coefficient for variable j
bi = right hand side value for constraint i
aij = coefficient associated with variable j in constraint i
We can show this portion of the initial simplex tableau as follows:

Table 2
(a)

(b)

c1

c2

...

cn

c row

a11

a12

ain

b1

a21

a22

a2n

b2

Matrix

Column

am1

am2

amn

bm

Thus for the Lowtech problem, we obtain the following partial initial simplex tableau
Table 3
x1

x2

s1

s2

s3

50

40

150

20

300

2.4

Improving the Solution

To improve the initial basic feasible solution, the simplex method must generate a

new basic feasible solution (extreme point) that yields a better value for the objective
functions. To do so require changing the set of basic variables and the simplex methods
provides an easy way to carry out this change of variables. For computational
convenience, we will add two new columns to the present form of the simplex tableau
and two rows to the bottom of the tableau.

Table 4
x1

x2

s1

s2

s3

Basis

CB

50

40

s1

150

s2

20

s3

300

zj

cj zj

50

40

Profit

Criterion for entering a new variable into the basis

Look at the evaluation row (cj zj) and select the variable to enter the basis that
will cause the larger per-unit improvement on the value of the objective function. In the
case of a tie, we follow the convention of selecting the variable to enter the basis that
corresponds to the leftmost f the columns.
Criterion for removing a variable from the current basis (minimum ratio test)
Suppose the incoming basic variable corresponds to column j in the A portion of
the simplex tableau, for each row I, compute the ratio b i/aij for each aij greater than zero.
The basic variable that will be removed from the basis corresponds to the minimum of
these ratios. In the case of a tie, we follow the convention of selecting the variable that
corresponds to the uppermost of the tied rows.

Table 5
Entering
Variable

Leaving
Variable

x1

x2

s1

s2

s3

bi/aij

Basis

CB

50

40

s1

150

150/3 = 50

s2

20

20/0 = (ignore)

s3

300

300/8 = 37.5

zj

cj zj

50

40

Adopting the usual programming terminology refer to this squared element as the pivot
element. The column and the row containing pivot element are called the pivot column
and pivot row respectively
Table 6
x1

x2

s1

s2

s3

Basis

CB

50

40

s1

25/8

-3/8

75/2

s2

20

x1

50

5/8

1/8

75/2

zj
cj zj

1875

Profit

Complete Second Tableau

Table 7
x1

x2

s1

s2

s3

Basis

CB

50

40

s1

25/8

-3/8

75/2

s2

20

x1

50

5/8

1/8

75/2

50

250/8 0

50/8

1875

70/8

-50/8

zj
cj zj

Moving Towards a better Solution

Table 8
x1

x2

s1

s2

s3

bi/aij

Basis

CB

50

40

s1

25/8

-3/8

75/2 / 25/8 = 12

s2

20

20/1 = 20

x1

50

5/8

1/8

75/2

75/2 / 5/8 = 60

zj

50

250/8 0

50/8

1875

cj zj

70/8

-50/8

The new tableau resulting from there row operations is as follows

Table 9
Basis
s1
s2
x1
zj
cj zj

CB
40
0
50

x1
50
0
0
1
50
0

x2
40
1
0
0
40
0

s1
0
8/25
8/25
-5/25
14/5
-14/5

s2
0
0
1
0
0
0

s3
0
-3/25
3/25
5/25
26/5
-26/5

12
8
30
1980

Note that the values of the basic variables are x2 = 12, s2 = 8, and x1 = 30, and the
corresponding profit is 40 (2) + 0 (8) + 50 (30) = 1980.
Optimally Criterion
The optimal solution to a LPP has been reached when all of the entries in the act
evaluation row (cj zj) are zero or negative. In such cases, the optimal solution is the
current basic feasible solution.
Interpreting the optimal solution
The optimal solution of the Lowtech problem, consisting of the basic variables x 1,
x2 and s2 and non basic variables s1 and s3 is written as follows.
x1 = 30, x2 = 12, s1 = 0, s2 = 8, s3 = 0.
Indicating that there are no idle unit of the assembly time constraint and the
warehouse capacity constraint, in other words, these constraints are both binding.
Moreover, if management wants to maximize total profit contribution, Lowtech should
produce 30 units of the Deskpro and 12 units of the portable. Since s2 = 8, management
should note that there will be eight unused portable display units.
SUMMARY OF THE SIMPLEX METHOD
The steps followed to solve a linear program using the simplex method can now be
summarized as follows:
We assume that the problem has all less than or equal to constraints and involves
maximization.
Step:
1.

Formulate a LP model of the problem

2.

Add slack variables to each constraints to obtain standard form

3.

Set up the initial simplex tableau

4.

Choose the non basic variable with the largest entry in the net evaluation row to
bring into the basis. This identifies the pivot column; the column associated with
the incoming variable.

5.

Choose as the pivot row that row with the smallest ratio of b i/ aij for aij > 0 where j
is the pivot column

6.

Perform the necessary elementary row operations to convert the column for the
incoming variable to a unit column with a 1 in the pivot row.

7.

Test for optimally. If cj zj 0 for all columns, we have the optimal solution. If
not, return to Step 4.

Problems
1.

Suppose a manufacturer produces to liquids, X and Y. Because of the past sales

experience the market researcher estimates that at least twice as much Y as X is
needed. The manufacturing capacity of the plant allows for a total of 9 units to be
manufactured. If each unit of liquid X results in a profit of N2 and the profit for
each unit of Y is N1, how much of each should be produced to maximize the
profit?

2.

Solve the following LP using the Simplex method.

Max 4x1 + 6x2 + 3x3 + x1
s.t
3/2x1 + 2x2 + 4x3 + 3x4 550
4x1 + x2 + 2x3 + x4 700
2x1 + 3x2 + x3 + 2x4 200
x1, x2, x3, x4 0.

3.

Basis

The following partial initial simplex tableau is given:

CB

x1

x2

x3

s1

s2

s3

20

25

40

30

-1/2

15

zj
cj zj
(a)

Complete the initial tableau

(b)

Write the problem in tableau form

(c)

What is the initial basis, does this correspond to the origin? Explain.

(d)

What is the value of the objective function at this initial solution?

(e)

For the next iteration, what variable should enter the basis, and what variable
should leave the basis?

2.5

Tableau Form: The General Case

(a)

Grater Than or Equal to Constraints

Suppose that in the Lowtech industry problem, management wanted to ensure that

the combined total production for both models would be at least 25 units. That enables us
to add another constraint to the constraint equation i.e. total production constraint. 1x 1 +
1x2 25
Max 50x1 + 40x2
s.t
3x1 + 5x2 150
1x2 20
8x1 + 5x2 300
1x1 + 1x2 25
x1, x2 0.
Standard Form:
Max 50x1 + 40x2 + 0s1 + 0s2 + 0s3 + 0s4
s.t
3x1 + 5x2 + s1 = 150
1x2 + s2 = 20
8x1 + 5x2 + s3 = 300
x1, x2, s1, s2, s3, s4 0
Tableau form:
3x1 + 5x2 + s1 = 150
1x2 + s2 = 20
8x1 + 5x2 + s3 = 300
1x1 + 1x2 1s4 + 1a4 = 25
We can write the objective function for the tableau form of the problem as follows:

Max 50x1 + 40x2 + 0s1 + 0s2 + 0s3 + 0s4 Ma4

Exercise: Prepare the initial simplex tableau for the problem. Hence obtain the optimal
solution
(b)

Equality Constraints
When an equality constraints occurs in a LPP, we need to add an artificial variable

to obtain tableau form and an initial basic feasible solution for example, if the equality
constraint is 6x1 + 4x2 5x3 = 30
With the artificial variable, the above equation becomes
6x1 + 4x2 5x3 + 1a1= 30
(c)

Negative Right-Hand Sides

Case 1: Equality constraints: 6x1 + 3x2 4x3 = -20

6x1 3x2 + 4x3 = 20
Case 2: Constraints:

6x1 + 3x2 4x3 -20

6x1 3x2 + 4x3 20

Case 3: Constraints:

6x1 + 3x2 4x3 -20

6x1 3x2 + 4x3 20

Example: Convert the following example problem into tableau form and then set up the
initial simplex tableau:
Max 6x1 + 3x2 + 4x3 + 1x4
s.t
-2x1 1/2x2 + 1x3 6x4 = -60
1x1 + 1x3 + 2/3x4 20
-1x2 5x3 - 50
x1, x2, x3, x4 0
We later obtain the tableau form:
Max 6x1 + 3x2 + 4x3 + 1x4 + 0s2 + 0s3 Ma1 Ma3
s.t.
2x1 + 1/2x2 1x3 + 6x4

1a1 = 60

1x1 + 1x3 + 2/3x4 + 1s2 20

1x2 + 5x3 1s3 + 1a3 50

x1, x2, x3, x4, s2, s3, a1, a3 0
The initial simplex tableau corresponding to this tableau form is:
Table 10
Basis
a1
s2
a3
zj
cj zj

CB
-M
0
-M

x1
6
2
1
0
-2M
6+2M

x2
3

0
1
-3/2M
3 + 3/2M

x3
4
-1
1
5
-4M
4 + 4M

x4
1
6
2/3
0
-6M
1 + 6M

s2
0
0
1
0
0
0

s3
0
0
0
-1
M
-M

a1
-M
1
0
0
-M
0

a3
-M
0
0
1
-M
0

60
20
50
-110M

Note that the squared element is the pivot element indicating that x4 will enter and a1 will
leave the basis at the first iteration.
2.6

Solving a Minimization Problem

There are two different ways of solving a minimization problem

(1)

Reverse the maximization rule i.e. select the variable with the most negative cj zj
as the one to introduce into solution. In this case, optimal solution is reached when
every value in the cj zj row is non-negative

(2)

Minimize z subject to a set of constraints should be changed to maximize (-z)

subject to a set of the same constraints.
i.e. min z = max (-z).

Example: Solve the LPP:

Min 2x1 + 3x2
s.t.
1x1 125
1x1 + 1x2 350
2x1 + 1x2 600
x1, x2 0.
Solution:
Convert to an equivalent maximization problem:
Max 2x1 3x2
s.t

1x1 125
1x1 + 1x2 250
2x1 + 1x2 600
x1, x2 0.
The tableau form for this problem is as follows:
Max -2x1 3x2 + 0s1 + 0s2 + 0s3 Ma1 Ma2
s.t
1x1 1s1 + 1a1 = 125
1x1 + 1x2 1s2 + 1a2 = 350
2x1 + 1x2 + 1s3 = 600
x1, x2, s1, s2, s3, a1, a2 0
The initial simplex tableau is shown below:
Table 11
x1
x2
s1
Basis CB
-2
-3
0
a1
-M
1
0
-1
s2
-M
1
1
0
a3
0
2
1
0
zj
-2M
-M
M
cj zj
-2+2M
-3+M
-M
At the first iteration, x1 is brought into the basis

s2
0
0
-1
0
M
-M
and a1

s3
a1
0
-M
0
1
0
0
1
0
0
-M
0
0
is eliminated.

a2
-M
0
125
1
350
0
600
-M
-475M
0
After dropping the

a1 column from the tableau, the result of the first iteration is as follows:
Table 12
x1
x2
s1
Basis CB
-2
-3
0
X1
-2
1
0
-1
a2
-M
0
1
1
s3
0
0
1
2
zj
-2
-M
2M
cj zj
0
-3 + M
-2+M
Continuing with two more iterations of the
tableau shown below:

s2
s3
a2
0
0
-M
0
0
0
125
-1 0
1
225
0
1
0
350
M 0
-M
-M 0
0
-250 225M
simplex method provides the final simplex

Table 13
x1
x2
s1
s2
s3
Basis
CB
-2
-3
0
0
0
x1
-2
1
0
0
1
1
250
x2
-3
0
1
0
-2
-1
100
s1
0
0
0
1
1
1
125
zj
-2
-M
2M M
0
cj zj
0
-3 + M
-2+M -M
0
-800
The simplex method provides the optimal solution with x1 = 250, x2 = 100, s1 = 125, s2 =
0, and s3 = 0. Note however that the value of the objective function is -800 in the final
simplex tableau. We must now multiply this value by -1 to obtain the value of the
objective function to the original minimization problem (total cost = 800).
2.7

Special Class

1.

Infeasibility: This occurs when an optimal solution is reach but one or more

artificial variables remained in the solution with a positive value. Degeneracy occurs
whenever there is no solution to the linear program that satisfies all the constraints,
including the non negativity constraints.
2.

Unboundedness: Occurs if the value of the solution may be made infinitely large

without violating any constraints. Unboundness indicates error in the problem

formulation. In this case, the simplex rule for determining the variable to be removed
from the solution will not work.
3.

Alternate Optimal Solutions: Exercise of two or more optimal solutions. This

occurs whenever the objective function line is parallel to one of the constraint lines. If the
linear programming problem has alternate optimal solution, cj zj will equal zero for one
or more of the variables nor in solution.
4.

Degeneracy: Occurs when one or more of the variables in the basic solution has a

value of zero. This does not create problem when it occurs at optimal solution. But if the
situation occurs at the early stage of iterations, an optimal solution may never be obtained
since each successive iteration will alternative between the same set of non optimal
points.

DUALITY THEORY AND SENSITIVITY ANALYSIS

Duality is an important concept in LP problem. Every LP problem has and
associated LP problem called the Dual. If the original problem known as the primal
problem, is a maximizing one then the dual formation is a minimizing one and vice versa.
A fundamental property of the primal-dual relationship is that the optimal solution to
either the primal or dual problem also provides that optimal solution to the other. In case
where the primal and dual problems differ in terms of computational difficulty, we can
choose the easier problem to solve.
Formulation of the Dual Problem
Consider the high tech industries problem described above:
Max 50x1 + 40x2
s.t.
3x1 + 5x2 150 assembly time
1x2 20 portable display
8x1 + 5x2 300 warehouse space
x1 , x 2 0
The High Tech dual problem is as follows:
Min 150u1 + 20u2 + 300u3
s.t.
3u1 + 8u3 50
5u1 + 1u2 + 5u3 40
u1 , u 2 0
The following general statements can be borne in mind about the dual problem.
1.

The dual is a minimization problem

2.

When the primal has n decision variables, the dual will have n constraints. The
first constraint of the dual is associated with variable x 1 in the primal and so on.

3.

When the primal has m constraints, the dual will have m decision variables. Dual
variable u is associated with the first primal constraint and soon.

4.

The r.h.s of the primal constraints becomes the objective function co-efficient in
the dual.

5.

The objective function co-efficient of the primal become the r.h.s. of the dual
constraints.
The constraint co-efficient of the ith primal variable becomes the co-efficient in the

6.

ith constraint of the dual.

The Optimal Dual Solution
Obtain the standard and tableau form as follows:
Max (-150u1 20u2 300u3 + 0s1 + 0s2 Ma1 Ma2)
s.t.
3u1 + 8u2 s1 + a1 50
5u1 + u2 + 5u3 s2 + a2 40
u1, u2, u3, s1, s2, a1, a2 0
The Initial Simplex Tableau is:
Table 14
Basis
a1
a2

CB
-M
-M

Zj
cj zj

U1
-150
3
5
-8M
-150 + M

U2
-20
0
1
-M
-20 + M

U3
-300
8
5
-13M
-300 + 13M

S1
0
-1
-1
M
-M

S2
0
0
-1
M
-M

a1
-M
1
0
-M
0

a2
-M
0
1
-M
0

At the first iteration, u3 is brought into the basis of a1 is removed.

The second tableau, with the a1 column dropped, is shown below.
Table 15
Basis
u3
a2

CB
300
-M

Zj
cj zj

U1
-150
3/8
25/8
(-900 25m)
8
(-300+25M)
8

U2
-20
0
1
-M

U3
-300
8
0
-300

-20+M

S1
0
-1/8
5/8
300 5M
8
-300+5M
8

S2
0
0
-1
M

a2
-M
0
1
-M

-M

50/8
70/8
-15,000 70M
8

50
40
-90M

At the second iteration u1 is brought into the basis, and a2 is removed. The third tableau,
with the a2 column removed is:
Table 16
Basis
u3
u1

CB
-300
-150

Zj
cj zj

U1
-150
0
1
-150
0

U2
-20
-3/25
8/25
-12
-8

U3
-300
1
0
-300
0

S1
0
5/25
5/25
30
30

S2
0
3/25
-8/25
12
-12

26/5 = 5.20
14/5 = 2.80
-1980

The optimal solution has been reached with u1 = 14/5, u2 = 0, u3 = 26/5, s1 = 0, s2 =

0. The value of the objective function for the optimal dual solution be (-1980) = 1980.
Observe that the original High Tech Industries problem yield also the same value
of the objective function. Thus, the optimal value of the objective function is the same for
both. This relationship is true for all optimal and dual LP problems.
Interpretation of Dual Prices
The dual price of a binding constraint provides valuable guidance because it
indicates to management, the extra contribution they would gain from increasing by one
unit the amount of scarce resource. As an example, the dual price of Assembly time is
N2.8 i.e. each unit of first resources contributes N2.80 in the objective function, while
each unit of 3rd resource contributes N5.20 in the objective function. This conclusion is of
course true so long as the current primal (dual) solution is optimal and feasible.
Using the dual to identify the primal solution
Recall the fundamental property of the primal-dual property, stated in the
preamble, also observe that when the primal problem is solve by the simplex method, the
optimal values of the primal variables appear in the right-most column of the final tableau
and the dual prices (values of the dual variables) are found in the zj row. Since the final
simplex tableau of the dial problem provides the optimal values of the dual variables, the
values of the primal variable should be found in the zj row of the optimal dual tableau.
This is infact the case and is formally stated as follows: Given the simplex tableau
corresponding to the optimal dual solution, the optimal values of the primal decision
variable are given by the Zj entries for the surplus variables; furthermore, the

optimal values of the primal slack variables are given by the negative of the Cj Zj
entries for the Uj variables.
This property enables us to use the final simplex tableau for the dual of the High
Tech problem to determine the optimal solution of x1 = 30 units of the Deskpro and x2 =
12 units of the portable. These optimal values of x 1 and x2 as well as the values for all
primal slack variables are given in the Zj and Cj Zj rows of the final simplex tableau of
the dual problem, which are shown again below.
Table 17
Basis
u3
u1

CB
-300
-150

Zj
cj zj

U1
-150
0
1
-150
0

U2
-20
-3/25
8/25
-12
-8

U3
-300
1
0
-300
0

S1
0
5/25
5/25
30
30

S2
0
3/25
-8/25
12
-12

26/5
14/5
-1980

Post-Optimal Analysis (Sensitivity Analysis)

After the optimal solution of a LPP has been reached, it is desired to study the
effect of discrete changes in the different co-efficient of the problem on the current
optimal solution. This study is referred to as sensitivity analysis. The usual sensitivity
analysis for LP programs involving computing ranges for the objective function coefficient and the right-hand-side value, as well as the dual prices.
In general, these changes may result in one of these cases:
1.

The optimal solution remains unchanged i.e. the basic variable and their value
remains essentially unchanged.

2.

The basic variables remain the same but their values are changed

3.

The basic solution changes completely.

Next, we illustrate sensitivity analysis with the objective function co-efficient.

Changes in the co-efficient of the Objective Function
Changes in the coefficients of the objective function can only affect the net
evaluation row cj zj equation and hence the optimality of the problem. In order to carry
out the sensitivity analysis, a range is placed on the coefficients value called the range of
optimality.

Recall the High Tech industries problem, the LPP is restarted below
Max 50x1 + 40x2
s.t.
3x1 + 5x2 150 assembly time
1x2 20 portable display
8x1 + 5x2 300 warehouse space
x1 , x 2 0
The final simplex tableau of the above LPP is reproduced below.
Table 18
X1

X2

S1

S2

S3

Basis

CB

50

40

X2

40

8/25

-3/25

12

S2

-8/25

3/25

X1

50

-5/25

5/25

30

Zj

50

40

14/5

26/5

1980

Cj z j

-14/5

-26/5

The range of optimality for an objective function coefficient, then, is determined by those
coefficient values that maintain
Cj zj 0 for every j

(*)

Thus, to determine the range of optimality for an objective function coefficient, say, ck,
we complete the value of the left-hand-side of inequality (*), using ck as the objective
function coefficient for xk.
Let us compute the range of quantity for c1, the profit contribution per unit of the
Deskpro using c1 (instead of 50) as the objective function coefficient of x 1, the revisd
final simplex tableau is as follows.

Table 19
X1

X2

S1

S2

S3

Basis

CB

50

40

X2

40

8/25

-3/25

12

S2

-8/25

3/25

X1

50

-5/25

5/25

30

C1

40

64 c1
5

C1 24
5

480 + 30c1

C1 64
5

24 c1
5

Zj
Cj z j

The current solution will remain optimal as long as the value of c1 results in Cj zj 0
for the two non-basic variables s1 and s3.
Hence we must have
c1 64 / 5 0 and 24 c1 / 5 0. Using these inequalities,
We obtain c1 64 or 24 c1

(**)

Since c1 must satisfy (**), the range of optimality for c 1 is given by 24 c1 64.
To verify this range, we recomputed the final simplex tableau after reducing the valueo f
c1 to 30.
Table 20
X1

X2

S1

S2

S3

Basis

CB

30

40

X2

40

8/25

S2

-8/25

3/25

X1

50

-5/25

5/25

30

Zj

30

40

34/5

6/5

1380

Cj z j

-34/5

-6/5

0 -3/25

12

Since cj zj 0 for all variable, the original solution is still optimal i.e. the optimal
solution with c1 = 30 is the same as the optimal solution with c1 = 50. Note, however, the
decrease in total profit from N1980 to N1380.
Exercise:
Try and investigate the range of optimality for c1 if the profit contribution per unit were
reduced even further say to N20
LINEAR PROGRAMMING: TRANSPORTATION PROBLEM
This section presents the transportation model and its variants. In the obvious
sense, the model deals with the determination of a minimum-cost plan for transporting a
single commodity from a number of sources (e.g. factories) to a number of destinations
(e.g. warehouses). The model can be applied to a wide variety of problems including
inventory control, employment scheduling, personnel assignment, scheduling dam
reservoir and many others.
It is basically a special type of linear programming problem that can as well be
solved by the regular simplex method. However, its special structure allows the
development of a solution procedure called the transportation technique i.e.
computationally more efficient. The model also can be modified to account for multiple
commodities.
The Transportation Model (definition and application)
The transportation model is formulated for a class problem with the following
unique characteristics:
1.

A product is transported from a number of sources to a number of destinations at

the minimum possible cost, and

2.

Each source is able to supply a fixed number of units of the product to each
destination, which has a fixed demand for the product.
Since there is only one commodity a destination can receive its demand from one

or more sources. The objective of the model is to determine the amount to be shipped
from each source to each destination such that the total transportation cost is minimized.

Let xij represents the amount transported from source i to destination j: then the LP
model representing the transportation problem is given generally as
Minimize z = cijxij
s.t.
xij ai

i = 1, 2, , m

Supply

xij bj

j = 1, 2, , n

Demand

Xij 0 for all i and j.

We use the following notations in the above LPM of the transportation problem:
i = index for origins: i = 1, 2, , m.
j = index for destinations: j = 1, 2, , n.
cij = cost per unit of shipping from origin i to destination j.
aij = supply or capacity in units at origin i.
bj = demand in units at destination j.
The first set of constraints stipulates that the sum of the shipments from a source
cannot exceed its supply; similarly, the second set requires that the sum of the shipments
to a destination must satisfy its demand.
The model described above implies that the total supply ai must at least equal
total demand bj. When the total supply equals total demand (ai = bj) the resulting
formulation is called a balanced transportation model. The model will be used to
demonstrate the solution of a transportation problem.
Solution of the Transportation Problem
In this section, we introduce the details for solving the transportation model. The
method uses the steps of the simplex method directly and differs only in the details of
implementing the optimal and feasibility conditions.
The Transportation Technique
The basic steps of the TT are:
Step 1: Determine a starting feasible solution

Step 2: Determine an entering variable from among the non basic variables. If all such
variables satisfy the optimality condition (of the simplex method), stop; otherwise go to
Step 3.
Step 3: Determine the leaving variable of the current basic solution; then find the new
basic solution. Return to Step 2.
Determination of the Starting Solution
Transportation model are solved within the context of a tableau. When the
transportation tableau is used an initial solution can be found by several alternative
methods, including the North West corner method, the minimum cell cost method (least
cost) and Vogels approximation method.
The steps of the North West corner method can be summarized as:
1.

Allocate as much as possible to the cell in the upper left hand corner subject to
the supply and demand constraints.

2.

Allocate as much as possible to the next adjacent feasible cell.

3.

Repeat step 2 until all rim requirement are met.

Also the steps of the least cost method can be summarized as

1.

Allocate as much as possible to the feasible cell with the minimum transportation
cost.

2.

Repeat step 1 until all rim requirements are met.

For the Vogels approximation method, the steps are:

1.

Determine the penalty cost for each row and column by subtracting the lowest cell
cost in the row or column from the next lowest cell cost in the same row or
column.

2.

Select the row or column with the highest penalty cost.

3.

Allocate as much as possible to the feasible cell with the lowest transportation cost
in the row or column having the highest penalty cost.

4.

Repeat steps 1, 2, and 3 until all rim requirements have been met.
Once the initial basic feasible solution has been determined by any of the previous

three methods, the next step is to solve the model for the optimal (i.e. minimum cost)

solution. There are two basic solution methods: the stepping stone method and the
modified distribution method (MODI) or method of multipliers. We will illustrate with
the following example.
Example 9
Wheat is harvested in the Midwest and stored in grain elevators in the three cities
Kansas City, Omaha and Desmonies. These grains elevators supply three mills that
produce flour, Chicago, St. Louis and Cincinnati. Grain is shipped to the mills in railroad
can such capable of holding one ton of wheat. Each grain elevator is able to supply the
following number of tons (i.e. railroad cars) of wheat to the mill on the monthly basis.

Table 21
(a)
Grain Warehouse

Supply

1.

Kansas City

100

2.

Omaha

175

3.

Desmoines

275
600 tons

Each mill demands the following tons of wheat per month:

(b)
Mill

Demand

A.

Kansas City

200

B.

Omaha

100

C.

Desmoines

300

600 tons
The cost of transporting one ton of wheat from each grain elevator (source) to each mill
(destination) differs according to the distance and rail system. These costs are:

Table 22
Mill
Grain Elevator

Chicago
A
6

St. Louis
B
8

Cincinnati
C
10

1.

Kansas City

2.

Omaha

11

11

3.

Desmoins

12

Determine how many tons of wheat to transport from each grain elevator to each mill on
a monthly basis in order to minimize the total cost of transportation.
Solution:
The linear programming model for the problem is formulated as follows
Minimize

Z = 6x1A + 8x1B + 10x1C + 7x2A + 11x2B + 11x2C + 4x3A + 5x3B

+ 12x3C

s.t.
x1A + x1B + x1C = 150
x2A + x2B + x2C = 175

supply constraints

x3A + x3B + x3C = 275

x1A + x1B + x1C = 200
x2A + x2B + x2C = 100

demand constraints

x3A + x3B + x3C = 200

xij 0
In this model, the decision variables, xij represent the number of tons of wheat
transported from each grain elevator, i (where i = 1, 2, 3) to each mill j (where j = A, B,
C).

Table 23
A
1

6
X1A

8
X1B

7
X2A

Demand

10
X1C

11
X2B

Supply

150
11

175

12

275

X2C
5

X3A

X3B

X3C

200

100

300

600

Determine a feasible solution: Using the northwest corner method

Table 24
The Initial NW corner allocation
A

B
6

C
8

Supply
10

1 150

150
7

11

11

175

12

275

3
Demand

200

100

300

600

Table 25
The Second NW corner allocation
A

Supply
10

1 150

150
7

11

11

175

12

275

2 50

3
Demand

200

100

300

600

Table 26
The Third NW corner allocation
A

B
6

C
8

Supply
10

1 15

150
7

2 50

11

11

175

12

275

100
4

3
Demand

200

100

300

600

Table 27
The Initial Solution from NWC is:
A

B
6

C
8

Supply
10

1 150

150
7

2 50

11
100

175

12

275

25
5

3
Demand

11

275
200

100

300

600

The transportation cost of the solution is computed by substituting the cell allocations
(i.e. the amounts transported)
x1A = 150, x2C = 25, x2A = 50, x3C = 275, x2B = 100
Z = 6x1A + 8x1B + 10x1C + 7x2A + 11x2B + 11x2C + 4x3A + 5x3B + 12x3C
Z = 6 (150) + 8 (0) + 10 (0) + 7 (50) + 11 (100) + 11 (25) + 4 (0) + 5 (0) + 12 (275).
= N5,925
Using the minimum cell cost method (least cost rule) and following the steps highlighted
above strictly.
Table 28
The Initial Solution minimum cell cost allocation
A

B
6

C
8

Supply
10

1 150

150
7

2 50

11
100

11

175

12

275

25
5

3 200
Demand

200

100

300

600

Table 29
The second minimum cell cost allocation
A

Supply
10

150
7

11

11

175

12

275

3 200
Demand

75

200

100

300

600

Table 30
The initial Solution from MCC
A

B
6

C
8

10

25
7

150
11

11

175

12

275

175
4

3 200
Demand

Supply

200

5
75
100

300

600

x3A = 200, x3C = 75, x2C = 175, x1B = 25, x1C = 125.
The total cost of this initial solution is N4,550 as compared to the N5,925 total
cost of the NW corner initial solution.
The third method for determining an initial solution, Vogels approximation
method (also called VAM), is based on the concept of penalty cost or regret.

Table 31
The VAM penalty cost
A

Supply

10

150
7

11

11

175
4

12

3
Demand

275
200

100

300
3

600
1

Table 32
The Initial VAM allocations
A

Supply

10

150
7

11

11

2 175

175
4

12

3
Demand

275
200
2

100

300
3

600
2

Table 33
The Second VAM allocation
A

Supply

10

150
7

11

11

2 175

175
4

12

3
Demand

275
200

100

300

600
2

Table 34:
The third VAM allocation
A

B
6

Supply

10

150
7

11

11

2 175

175
4

3 25
Demand

200

12

100
100

275
300

600
2

Table 35
The Initial VAM solution
A

Supply
10

150
7

11

150
11

2 175

175
4

3 25
Demand

200

12

100

150

275

100

300

600

x1C = 150, x2A = 175, x3A = 25, x3B = 100, x3C = 150
The total cost of this initial solution is N5,125 which is not high as the N5,925
obtained by the NWC initial solution or as low as the MCC solution of N4,550. Like the
MCC method, VAM typically results in a lower cost for the initial solution than the NWC
method.
Step 2: Determination of Optimal Solution (Modified Distribution Method [MODI]
or Method of Multipliers)
In order to demonstrate MODI, we will again use the initial solution obtained by
the minimum cell cost method.
Table 36
The minimum cell cost initial solution
V A = VB = VC =
From

To
U1

B
6

Demand

10
150

11

175

11
175

4
U3

Supply

1
7

U2

200

75

200

100

175
12
275

300

600

The extra left hand column with the u1 symbols and the extra row with the vj symbols
represents column and row values that must be computed in MODI. These values are
computed for all cells with allocation by using the following formula: u i + vj = cij
The value of cij is the transportation cost for cell ij. For example, the formula for cell 1B
is u1 + vB = C1B and since CIB = 8
u1 + vB = 8
The formulas for the remaining presently allocated cells are:
x1C: u1 + vc = 10
x2C: u2 + vc = 11
x3A: u3 + vA = 4
x3B: u3 + vB = 5
Thus, if we let u1 = 0, then we can solve for all remaining u i and vj values
vB = 8
x1C: u1 + vc = 10
vc = 10
x2C: u2 + vc = 11
u2 = 1
x3A: u3 + vA = 4
vA = 7
x3B: u3 + vB = 5
u3 = -3
Now, all the ui and vj values can be substituted into our table as shown in Table 2

Table 37
The Initial Solution with all ui and vj values
VA = 7

VB = 8

From

VC = 10

To
U1 = 0

6
1

C
-

25

Demand

11

175

150
11

175
-

U3 = -3

10
125

7
U2 = 1

Supply

200

75

200

100

175
12
275

300

600

Next we use the following formula to evaluate all empty cells (non basic variables) cij
ui vj = kij
Where kij equals the cost increase or decrease that would occur by allocating to a cell.
For the empty cells in (Table 2)
x1A : k1A = c1A

u1 vA = 6 0 7 = -1

x2A : k2A = c2A

u2 vA = 7 1 7 = -1

x2B : k2B = c2B

u2 vB = 11 1 8 = +2

x3C : k3C = c3C

u3 vC = 12 (-3 10) = + 5

This indicates that either cell 1A or 2A will decrease cost by N1 per allocated ton.
We can select either cell 1A or 2A to allocate since they are tied at -1. If cell 1A is
selected as the entering non basic variable, then the stepping stone path (loop) for the cell
must be determine so that we know how much to reallocate.

Table 38
The second iteration of the MOD solution method
V A = VB = VC =
From

To
U1 =

B
+

1 25

U3 =

150
11

175
4

3 175

100

200

100

Demand

10

11

2 175
-

Supply

125
7

U2 =

175
12
275

300

600

The ui and vj values for (table 3) must now be recomputed using our formula for the
allocated cells
Let u1 = 0
x1A : u1 + vA = 6

vA = 6

x1C: u1 + vC = 10

vC = 10

x2C : u2 + vC = 11

u2 = 1

x3A : u3 + vA = 4

u3 = -2

x3b : u3 + vB = 4

vB = 7

Table 39
The new ui and vj values for second iteration
VA = 6

VB = 7

From

VC = 10

To

U1 = 0

1 25

Supply
10
150

125
7

U2 = 1

11

2 175

11
175

U3 = 2

3 175

100

Demand

200

100

175
12
275

300

600

The cost changes for the empty cells are not computed using the formula
cij ui vj = kij
x1B : k1B = c1B

u1 vB = 8 0 7 = -1

x2A : k2A = c2A

u2 vA = 7 1 6 = 0

x2B : k2B = c2B

u2 vB = 11 1 7 = +3

x3C : k3C = c3C

u3 vC = 12 (-2) 10 = + 4

Since none of these values is negative, the solution shown above in (table 39) is optimal.
Thus the optimal solution is calculated as: 6 (25) + 10 (125) + 11 (175) + 4 (175) + 5
(100) = N4,525
Steps of MODI (Methods of Multipliers)
(1)

Develop an initial solution using one of the three methods available

(2)

Compute u1 and vj values for each row and column by applying the formula ui vj
= cij to each cell that has an allocation

(3)

Compute the cost change, kij, for each allocated cell using the formula c ij ui vj
= kij.

(4)

Allocate as much as possible to the empty cell that will result in the greatest net
decrease in cost (kij). Allocate according to the stepping stone path or the selected
cell.

(5)

Repeat steps 2 through 4 until all kij values are positive or zero.

DEGENERACY
In the entire tableau showing solution to our wheat transportation problem, the
following condition was met. m row + n columns 1 = no of cells with allocation.
For example, in any of the balanced tableau for wheat transportation, the no of
rows are 2 (i.e. m = 3) and the no of columns are 3 (i.e. n = 3), thus:
3 + 3 1 = 5 cells with allocations.
Five cells with allocation always existed for these tableaus thus our condition for
normal solution was met. When this condition is not met and less then m + n 1 cells
have allocations, the tableau is said to be degenerate.
THE ASSIGNMENT MODEL
Consider the situation of assigning m jobs (or workers) to n machines. A job i (=
1, 2, , m) when assigned to machine j (= 1, 2, , n) incurs a cost c ij. The objective is to
assign the jobs to the machines (one job per machine) at the least total cost. The situation
is known as the assignment problem.
Table 40
machine
1

c11

c12

c1n

c21

c22

c2n

cm1

cm2

cmn

The formulation of this problem may be regarded as a special case of the

transportation model. Here jobs represents sources and machines represent
destinations. The supply available at each sources is 1; i.e. a i = 1 for all i, similarly, the
demand required at each destination is 1; i.e. bi = 1 for all j. the cost of (assigning) job i
to machine j is cij.
The mode is thus given by:
Minimize z: cijxij
Subject to
xij = 1,

i = 1, 2, , m

xij = 1,

j = 1, 2, , n

xij = 0 or 1
0

if the jth job is not assigned to the ith machine

otherwise

xij =

The Steps of the assignments solution method can be summarized as:

(1)

Perform row reduction by subtracting the minimum value in each row from all
other row values.

(2)

Person column reductions by subtracting the minimum value in each column from
all other column values

(3)

In the completed opportunity cost table, cross out all zeros using the minimum
number of horizontal and/or vertical lines

(4)

If less than m lines are required (where m = the no of rows or columns), subtract
the minimum uncrossed value from all other uncrossed values, and add this same
minimum value to all cells where two lines intersect. All other values are
unchanged.

(5)

If m lines are required, the optimal solution exists and m unique assignments are
made. If less m lines are required, repeat step 4.

Example 1: Court Scheduling

A court administrator is in the process of scheduling four court dockets. Four

judges are available to be assigned, one judge to each docket. The court administrator has
information regarding the types of cases on the dockets as well as data indicating the
relative efficiency of each of the judges in processing different types of court cases.
Based on this information, the court administrator has compiled the data in the Table
below:
Table 41
Docket
Judge

14

13

17

14

16

15

16

15

18

14

20

17

20

13

15

18

The table shows the estimates of the no of court days each judge would require in
order to completely process each court docket. The court administrator would like to
assign the four judges so as to minimize the total no of court days needed to process all
four dockets
Table 42
Docket
Judge

Table 43
Docket
Judge

Table 44
Docket
Judge

Table 45
Docket
Judge

Final Assignment
Judge 1 to Docket 1

14

Judge 2 to Docket 4

15

Judge 3 to Docket 2

14

Judge 4 to Docket 3

15

Total days

58

Thus the optimal assignment of the four judges results in 58 judges days to clear the
four dockets.

The Assignment Technique for Maximizing

A maximizing assignment problem typically involves making assignments so as to
maximize contribution. To maximize only step 1 from the above differs, the column are
reduced by the largest number of each column. From then on the same rules apply that
are used for minimizing.
Example: Suppose the figures below relate to contribution and that it is required to
maximize contribution.
Table 46
W

25

18

23

14

38

15

53

23

15

17

41

30

26

28

36

29

Table 47
Column
W

13

10

30

16

13

23

11

12

12

17

Table 46
~ Row
W

20

13

23

11

12

12

17

Table 49
~
W

20

13

23

11

12

12

17

Table 50
~
W

17

16

10

20

11

14

Table 51
~
W

17

16

10

20

11

14

Table 52
~
W

17

16

10

20

11

14

The assignments
C to Z

20

D to X

28

A to W

25

B to Y

53
136
Unequal sources and destinations
To solve assignment problems in the manner described the matrix must be square

i.e. the supply must equal requirement. Where the supply and requirements are not equal,
an artificial source or destination must be created to square the matrix. The costs/mileage
/ contribution etc. for the fictitious row or column will be zero throughout.

Solution method, having mad the sources equal to the destination, the solution
method will be as normal, treating the fictitious elements as though they are real. The
solution method will automatically assign a source or destination to the fictitious row or
column and the resulting assignment will incur zero or gain zero contribution.
Example:
A foreman has four fitters and has been asked to deal with five jobs. The times or
each job are estimated as follows:
Table 53
Fitters
Job

Alf

Bill

Charlie

Dave

12

20

12

22

18

15

20

12

16

18

15

16

12

20

18

14

10

17

Dummy fitter inserted to square the matrix

Table 54
A

Dummy

12

20

12

22

18

15

20

12

16

18

15

16

12

20

18

14

10

17

Table 55
A

Dummy

15

10

16

10

16

10

12

Table 56
A

Dummy

10

13

10

12

Table 57
A

Dummy

10

13

13

10

12

5 lines so optimum assignments:

B to 4
C to 5
A to 2
Dummy to 2
D to 3
Hence Job 2 is not done

NETWORK ANALYSIS/PROJECT MANAGEMENT

1.

INTRODUCTION TO NETWORK ANALYSIS

Definition
Network analysis is a generic term for a family or group of related techniques
developed to aid management to plan and control project.

Why Network Analysis?

It is most valuable when project are complex, large and restricted i.e. completed
within stipulated or cost limits. It illustrates the way in which parts of the project are
organized and determine the time duration of these projects. It also aids planning and
scheduling of projects.

Background
In the 1950s, a basic form of Network analysis was being used in the UK and
USA in order to reduce project times i.e. the amount of time spent to complete a project.
In 1950, the US Naval Special Projects Office set up a team to devise a means of
controlling the planning of complex projects. The team came up with a network
technique known as PERT i.e. Programme Evaluation and Review Technique. This
technique was used in planning and controlling the development of the Polaris missile
and credited with saving two years in the development of missiles.
Since 1958, the technique has been improved upon and nowadays many variants
exist which handle, in addition to basic time, costs, resources, probabilities and
combination of these factors. Variety of names exists and among the commonly used are:
CPP (Critical Path Planning), CPA (Critical Path Analysis), CPM (Critical Path Method),
PERT etc.

Network Techniques (CPM/PERT)

The network techniques that are used for project analysis are CPM (Critical Path
Method) and PERT (Programme Evaluation Review Techniques). They were developed

at approximately the same time, although independently during the late 1950s. The fact
that they have been frequently and widely used in such a short of period of time attests to
their value as management science techniques.
The basic difference between them is that CPM is a deterministic technique while
PERT is a probabilistic technique.
2.

BASIC NETWORK TERMINOLOGY

i.

Activity: This is a task or job which takes time and resources e.g. Building of
house, construction of bridges etc. It represented in a network by an arrow () in
which the head indicates the end of an activity and the tail indicates the beginning
of an activities.

ii.

Event: It is a point in time and it indicates the start and finish of an activity. It is
represented in a network by a circle or nodes (O). Note that the establishment of
activities automatically determines the event.

iii.

Dummy Activities: This is an activity which does not take time or resources. It is
used merely to show logical dependences or sequences between activities so as not
violate the rules of drawing network. It is represented in a network by a dotted
arrow (

iv.

Predessor Activity/Successor Activity: This is an activity that precedes/succeeds

an activity in a network or an activity that must be completed before the next
activity can start (predecessor activity).
An activity that can start after the one before it has been completed (successor
activity).
E.g

Design a house

Obtain a finance

Build House

From example, design a house is a predecessor activity for obtain finance

while building house is a successor activity for obtain finance.
v.

Network: This is the combination of activities, events, dummy activities in a

logical sequence according to the rules of drawing network.

Example of a network;

3.

RULES OF NETWORK CONSTRUCTION

1.

A complete network must have only one point of entry (start event) and only one
point of exit (finish event).

2.

Every activities must have one preceeding event (tail event) and one succeeding
event (head event).

Note: Many activities may use the same tail event or the same head event. E.g.

Tail event

Head Event

i.e. they cant use the same tail and head event.
3.

No activity can start until its tail event is reached

4.

An Event is not complete until all activities leading into it are complete

5.

Loops are not allowed i.e. series of activities which leads back to the same event
e.g

6.

All activities must be tied into a network i.e. danglers is not allowed

Note: Danglers are the activities that do not link into the overall network. e.g

Dangling activities

4.

CONVENTION FOR DRAWING NETWORKS

1.

Networks proceeds from left to right.

2.

Networks are not drawn to scale but must be drawn neatly.

3.

Arrows need to be drawn in the horizontal plane unless totally unavoidable, they
must proceed from left to right.

4.

Events should be progressively numbered from left to right.

Activity Identification
Activities can be identified in several ways, typical of the method to be found
includes:
1.

Shortened description of jobs e.g order materials

2.

Alphabetic or numeric code e.g A,B,C or 100,130 etc.

Dummy Activities
Dummy activities are the one that does not consume time or resources but shoe
logical relationship in a network. It is shown on a network by a dotted arrow.
Example: Assume that a car arrive at a service station during which two independent
activities takes place, filling with petrol (A) and topping up with oil (B).
This could be
Car
Arrived

Car
Ready

Which is incorrect because it contradicts the rule that says two or more activities cant
have the same head and tail event?
The correct network is:
Filled
with
Petrol

Car
Arrived

Car
Ready

Example 1: Draw the network of the following problem:

Activity

Preceding Activities

1 2, 3, 4

6 and 2

7 and 8

10

11

12

9, 10, 11

Solution:
5
7
8

6
1

10

12

11

Summary
1.

Network analysis is used for the planning and controlling of large complex
projects.

2.

Network comprises activities which consumes time and resources and also event
which is point in time. Activities is represented by an arrow ( ) while event by a
circle (O) in a network.

3.

Network has one start and one head event. An event is not complete if the activity
leading to it is not complete.

4.

Two or more activities cant have the same head and tail event.

5.

The length of the arrows representing the activities is not important because
networks are not drawn to scale.

6.

Dummy activities (represented by

) are necessary to show logical relationship.

They do not consume time or resources. They become necessary as the network is
drawn.
EXERCISE:
1.

Draw the network of the following:

12

23

46

78

13

36

57

80

25

45

58

24

47

67

12

45

79

13

48

14

57

25

58

26

87

35

69

2.

Solution to Exercise
1.

1
4

2.
2

5.

TIME ANALYSIS IN PROJECT SCHEDULING

Critical Path Calculations

The critical path of network gives the shortest time in which the whole project can
be completed. It is the chain of activities with the longest duration time.
It can be calculated using Inspection/Tree diagram method or Forward/Backward
pass method.
Inspection/Tree Diagram Method
This is done by writing out all the paths in a network then the path with the longest
duration time is the critical path.
Example 2

3
Lay
Foundation

3
Design
house

Dummy
0

2
1
Order
Material

Build
House

4
Select
Paint

3
1

Select
Carpet

Finish
Work
1

The paths are:

1.

123467

Duration times 3 + 2 + 0 + 3 + 1 = 9
2.

3
2
0 1
1 1
1234567

Duration times: 3 + 2 + 0 + 1 + 1 + 1 = 8
3.

124567

Duration times: 3 + 1 + 1 + 1 + 1 = 7
4.

12467

Duration times: 3 + 1 + 3 + 1 = 8
The longest duration is 9 months
Therefore, the critical path is 123467
5.2

Forward/Backward Pass Method

1.

Earliest Start Time, which is also the forward pass. This is the earliest possible

time at which a succeeding activity can start. The Formula is

ETj = max (ETi + tij) where i is the starting node number of all activities ending at node j.
tij is the time for activity i j
2.

Latest Start Time, which is also the backward pass. This is the latest possible

time at which a preceding activities can be completed without delaying beyond project
duration time. The formula is:
LTi = max (LTi tij) where j is the ending node number of activities starting at i t ij is the
time for activity i j.

CRITICAL ACTIVITIES
This is the activities that have the same earliest start time and latest start time.

Critical Path (Estimation of project completion time)

The path that has the same earliest start time and latest start time is called the
critical path which also corresponds to the chain of activities with the longest duration
time. It is indicated in a network by double line or different colour in the network.
Example 3: Use backward/forward method to calculate critical path of example 2
3

1
ET = 0
LT = 0

ET = 5
LT = 5

ET = 5
LT = 5

ET = 3
LT = 3

ET = 8
LT = 8

3
1

7
1

ET = 9
LT = 9

Project
duration time

ET = 6
LT = 7

For the earliest start time: ETj = max (ETi + tij)

ET1 = 0,

ET2 = ET1 + t12 = 0 + 3 = 3

ET3 = (ET2 + t23) = 3 + 2 = 5
ET4 = max (ET2 + t24, ET3 + t34) = max (3 + 1, 5 + 0) = 5
ET5 = ET4 + t45 = 5 + 1 = 6
ET6 = max (ET4 + t46, ET5 + t56) = max (5 + 3, 6 + 1) = 8
ET7 = ET6 + t67 = 8 + 1 = 0

For latest start time: LTi = min (LT j tij)

LT7 = 9, LT6 = LT7 + t67 = 9 1 = 8
LT5 = LT6 t56 = 8 1 = 7
LT4 = min (LT5 t45, LT6 t46) = min (7 1, 8 3) = 5
LT3 = LT4 t34 = 5 0 = 5
LT2 = min (LT3 t23, LT4 t24) = min (5 2, 5 1) = 3
LT1 = LT2 t12 = 3 3 = 0
Therefore, the critical path is 123467 and the project duration time = 9
months.

Slack of an Activity
This is the time an activity can be delayed without affecting the overall project
duration i.e. extra time available for completing an activities
Slack for activity i j = LTj ETj tij
Where LTj = Latest start time for activity j
ETi = Earliest start time for activity i
tij = Activity time between i and j
Slack of an event
This is difference between the earliest start time and the latest start time of each event.
Note: The activities or event on the critical path does not have activity slack or event
slack.
From the example above, the slack activity for 24 = 5 3 1 = 1 months i.e. there is
extra one month. The slack event for (5); = LT 5 ET5 = 7 6 = 1 month
Note: Events is the only event not on the critical path.

6.

PROBABILITY CONSIDERATIONS IN PROJECT SCHEDULING

PERT Analysis
In CPM, all the activities time estimate has a single value i.e. we are assuming that
the activity time estimate is known with certainty, but in reality, it is rare that activity
time estimate can be known with certainty. This is true since projects that are networked
tends to be unique as such there is little historical evidence to be used as basis to predict
future occurrence. As an alternative to CPM, PERT (Project Evaluation and Review
Technique) uses probabilistic activity times.
Example 4
Using the example below to demonstrate PERT
A Southern Textile Company has decided to install a new computerized orderprocessing system. In the past, order for the cloth the company produce were processed
manually, which contributes to delay in delivering orders and as a result, lost sales. The

company wants to know how long it will take to install the new system. The network is
given below:

Equipment resting and changes

System
Development

Manual Testing

Company Orientation

System Training

System
Changeover

All this state in CPM network, we should assign a single-time estimate to each
network activity. However, in PERT networks we will determine three time estimates for
each activity which will enable us to estimate the mean and variance for a beta
distribution of the activity times.
We assume that the activity time estimate can be described by beta distribution for
several reasons:
1.

The beta distribution mean and variance can be approximated by three estimates

2.

The beta distribution is continuous and has no predetermined shape it will take on
the shape that is indicated by the time estimate given.

3.

It has become traditional to use beta distribution for PERT analysis

The three estimates are (1) most likely time (m) (2) optimistic time (a) (3) pessimistic
time (b)
1.

Most Likely Time (m): Is the time would most frequently occur if the activities
were repeated many times

2.

Optimistic Time (a): Is the shortest possible time that an activity could be
completed assuming everything went right

3.

Pessimistic Time (b): Is the longest possible time required for an activity to
complete if everything went wrong.

These 3 estimates will be determined the mean and variance of beta distribution:

Mean (Expected time) =

2

Variance =
The three estimates mean and variance of the southern textile network is given below:
Time estimate (week)
a
m
b
6
8
10
3
6
9
1
3
5
0
0
0
2
4
12
2
3
4
3
4
5
2
2
2
3
7
11
2
4
6
0
0
0
1
4
7
1
10
13

Activities
12
13
14
25
26
35
45
48
57
58
87
69
79

mean time
t
8
6
3
0
5
3
4
2
7
4
0
4
9

Variance
4/9
1
4/9
0
25/9
1/9
1/9
0
16/9
4/9
0
1
1

To calculate the mean for activity 1 2, a = 6, m = 8, b = 10

t==

=8
2

Variance =

Calculate the critical path using EST and LST

ET = 9
LT = 9

ET = 13
LT = 21

t=5

t=6

t=4

t=0

t=8
ET = 6
LT = 6

ET = 9
LT = 9

t=7

ET = 0
LT = 0
t=4

t=3

4
ET = 3
LT = 5

t=4

t=2

t=0

ET = 13
LT = 16

t=9
ET = 16
LT = 16

ET = 25
LT = 25

ETj = max (ETi + tij)

LTi = min (LTj tij)
The critical path is 13579 and the project duration = 25 weeks
The project duration or expected time and the variance can be computed by
summing the, expected time and variance for each activities in the critical path because
PERT method assumes that activities are statistically independent
Variance activities

Variance

13

35

1/9

57

16/9

79

1 + 1/9 + 16/9 + 4

= 62/9 weeks
= 6.9 weeks

PERT also assumes that mean and variance of the network is normally distributed
7.

ESTIMATION ON PROJECT COMPLETION TIME USING STANDARD

NORMAL VARIANCE VALUE

We can interpret the expected time and variance as mean and variance as mean and
variable of normal distribution.
From above, mean = 25 weeks
Variance 2 = 6.9 weeks
Normal Distribution

z=

where is standard deviation

x proposed project completion time

Example 5: Suppose the textile company manager told customers the new orderprocessing system would be completely installed in 30 weeks. What is the probability
that it will in fact be ready by that time?

Solution:
x = 30 weeks, = 25 weeks, 2 = 6.9 weeks = 2.63

= 25

x = 30

z=

= 1.90

The z value 1.90 corresponds to 0.4713 on the normal distribution table, this
means there is a possibility 0.5 + 0.4713 = 0.9713 of completing the project in 30 weeks.
Alternatively, suppose one customer is so frustrated with delay order has told the textile
company that if she does not have the new ordering system working in 22 weeks, she will
trade elsewhere.
Solution:

x = 22

z=

The value on the normal table corresponds to 0.3729. This mean the probability of
completing the work in 22 weeks is 0.5 0.3729. 0.1271.
Summary
1.

Basic time analysis is the calculation of critical path which is the shortest time
which the project is completed.

2.

To determine the critical path, calculate the EST and LST for each event and
compare them. The chain activities in which the LST and EST are the same is the
critical path.

3.

The event or activities on the critical path has zero slack

4.

Given the 3 time estimates: Optimistic, most likely time, pessimistic the project
time and variance can be computed.

5.

The probability of the project time is calculated using =

Exercises
a.

Find the critical path of the following network using the EST/LST.
Activity

Preceeding Activity

Duration (Days)

2,6

11

7.8

10

11

12

9, 10, 11

b.

Find the activity slack and event slack in the network

2.

Assuming the variance of the activities on the critical path in question 1 are 1, 4,

2.25, 9, 6.25, 9 respectively. Based on this value calculated the responsibility of

achieving a schedule time of 40 days for the project duration
ET = 18
LT = 13

ET = 16
LT = 11

2
2
A
ET =
0

1
4

B
ET =
4

F
3

ET = 23
LT = 23

9
ET = 12
LT = 12

7
4

10

ET = 9
LT = 9

ET = 30
LT = 30

11

3
H

ET = 10
LT = 27

12

J
4

ET = 34
LT = 34

The double line indicates he critical path

b.

Activity slack: LTj ETi tij

B C = LTC ETB tBC = 16 4 7 = 5 days
C E = LTE ETC - tCE = 18 11 2 = 5 days
E G = 23 13 5 = 5 days
D I = 30 9 4 = 17 days
B H = 27 4 6 = 17 days
H I = 30 10 3 = 17 days

Slack Event:
Event C = 16 11 = 5 days
Event E = 18 13 = 5 days
Event H = 22 10 = 17 days
Variance = 1 + 4 + 2.25 + 9 + 6.25 + 9
2 = 31.5

INTEGER PROGRAMMING
In this section, we focus on a class of problems that are modeled as linear
programs with the additional requirement that some requirement that some or all of the
decision variable must be integer. Such problems are called Integer Linear
Programming Problems. The use of integer variables provides additional modeling
flexibility. As a result, the number of practical applications that can be addressed with LP
methodology is enlarged and includes capital budgeting, distribution system Design,
Location problems etc.

Types of integer Linear Programming Models

Formally, the general integer programming problem is
Maximize f(x)
s.t
gj (x) = 0,

j = 1, 2, , m

hi (x) 0,

i = 1, 2, , k

X = (x1, x2, , xq, xq+1 , , xn)

Where x1, x2, , xq are integers for a given q. As problem remains essentially
unsolved in the general case we confine our attentive to a useful simplification. We
assume f and the his are linear, these are no gjs and all the variables in X must be nonnegative. Then the formulation can be expressed in matrix notation as
Maximize CX

(1)

AX b

(2)

X0

(3)

x1, x 2, , xq , integers

(4)

s.t

Where X = (x1, x2 , , xq +1, , xn)T, C is a 1 x n real vector, b is an m x 1 real vector, A

is an m x n real matrix.
If q = n, the problem is termed an all- integer linear programming problem

If 1 < q < n, the problem is termed a mixed- integer linear programming problem. If
(x1, x2, , xq) is replaced by xj = 0 or 1, i =1, 2..n problem is termed a zero-one
programmed problem.

Rounding (Graphical Solution)

One obvious approach to (1) (4) is to neglect (4) and solve the resulting problem
graphically. If the solution produced satisfies (4) then it must be optimal. If it does not
then there are a number of options available. One straight forward strategy is to round the
values of non integer values either up or down to achieve n integer solution.
Let us explore this idea on the following integer programming problem:
(1)

Max z = 2x1 + 3x2

s.t
195x1 + 273x2 1365
4x1 + 40x2 140
x1 4
x1 x2 0 and integer

The Linear programming version of this problem has been solved graphically in (fig 1.0)
it can be been that the optimal solution is
(x1, x2) = (2.44, 3.26)

Z = 14.66

Rounding the decision variable to the nearest integer value yields a solution of x 1 =2 and
x2 = 3 for an objective function value of 13 or #13,000 annual cash flow. In fig 1, we
shows the feasible solution points that provide integer values for x 1 and x2. Is the rounded
solution (x1, x2) = (2, 3) the optimal solution? The answer is no! as can be seen that the
optimal integer solution is x1 = 4 and x2 = 2, with object function value of 14.00 or
#14,000 annual cash flow.

Rounded LP
Solution (2,
3)

4
Opt. LP Relaxation
Solution (2.44,
3.26)
3

Opt. Integer
Solution (4, 2)

(3, 2)
Feasible
region for the
LP relaxation
(2, 2)

Fig (1.0)
Graphical Solution to the LP Relaxation

Methods of Integer Programming

The section unveils the methods that guarantees to find an optimal solution (if one
exists) to any integer programming problem. The two broad approaches Branch and
Bound Technique and Cutting plane method. The earlier technique starts with the
continuous optimum; but systematically partitions the solution space into sub problems
by deleting parts that contain no feasible integer points. The cutting methods
systematically adding special secondary constraints, which essentially represent
necessary conditions for integrality, the continuous solution space is gradually modified
until its continuous optimum extreme points satisfies the integer conditions.

The branch and bound algorithm was originally developed by A.H. Larid and A.
G. Doig. However, R.J. Dakins modification offers greater computational advantage and
his version will be presented here.
Branch and Bond Solution
BB is currently the most efficient general-purpose solution procedure or integer
linear programs. The BB procedure begins by solving the LP Relaxation of the integer
linear program. The LP Relaxation of the above problem is stated below,
Max 2x1 + 3x2

x1 = blocks of town houses purchased

s.t

x2 = apartment building

195x1 + 273x2 1365

funds available

4x1 + 40x2 140

Managers time

x1 4

Town house availability

x1 , x 2 0
UB = 14.66 (X1 = 2, X2 =
3
LB = 13.00 Value = 13)

Max 2x1 + 3x2

s.t
195x1 + 273x2 1365
4x1 + 40x2 140
x1 4 x1 2
x1, x2, 0

Note:

1
LPR Value
=14.66
X1 = 2.44
X2 = 3.26

x1 3

x1 2

2
LPR Value
=13.90
X1 = 2
X2 = 3.30

UB = 14.00
LB = 14.00

Best feasible integer solution

found: x1 = 4, x2 = 2; value =
14 see node 4,

Max 2x1 + 3x2

s.t
195x1 + 273x2 1365
4x1 + 40x2 140
x1 4 x1 3
x1, x2, 0

3
LPR Value
=14.58
X1 = 3
X2 = 2.86

x2 3
Max 2x1 + 3x2
s.t
195x1 + 273x2 1365
4x1 + 40x2 140
x1 4 x1 2
x1, x2, 0

x2 2

4
LPR Value
=14.00
X1 = 4
X2 = 2

Complete BB Solution

Max 2x1 + 3x2

s.t
195x1 + 273x2 1365
4x1 + 40x2 140
x1 4 x1 3
x1, x2, 0

5
Infeasible

The Branch and bound solution procedure could be summarized as follows

1.

Solve the LP Relaxation of the IP at node 1 set UB value equal to the value of the
LP Solution

2.

Find n feasible integer solution. Set LB equal to the valueo f the feasible integer
solution

3.

Is UB = LB? If yes, the optimal solution is the feasible solution with value = LB

4.

Otherwise branch from the node with the greatest LP value. Find the variable
(call it xj) that is furthest from being integral. Create two branches and two
descendant nodes; one with xj (xj) and one with xj (kj) + 1.

5.

Solve the LP Relaxation at each of the descendant nodes, and record its LP value

6.

Re compute the upper bound by finding the maximum over all nod from which
there are no branches

7.

Re compute the lower bound as the max value of all feasible integer solution
found to date. Test for 93) and treat the decision arrive at accordingly.

Extension to Mixed-Integer Integer Programs

One of the advantage of BB Solution procedure for integer programming is that it
is applicable to both all-integer and mixed-integer linear programs. To see how the BB
solution approach can be applied to a mixed-integer linear program, let us return to
problem (1) and suppose that x2 was not required to be integer i.e.
Max 2x1 + 3x2
s.t
195x1 + 273x2 1365
4x1 + 40x2 140
x1 4
x1, x2, 0 and x1 integer
The BB solution procedure is illustrated using the decision tree below:

Note:

Max 2x1 + 3x2

s.t
195x1 + 273x2 1365
4x1 + 40x2 140
x1 4 x1 2

1
LPR Value
=14.66
X1 = 2.44
X2 = 3.26

x1 3

x1 2

x1, x2, 0 and x1 integer

2
LPR Value
=13.90
X1 = 2
X2 = 3.30

Compare
UB = 14.58

UB = 14.58
LB = 14.58

Best feasible integer solution

found: x1 = 3, x2 = 2.86;
value = 14.58

Max 2x1 + 3x2

s.t
195x1 + 273x2 1365
4x1 + 40x2 140
x1 4 x1 3
x1, x2, 0

3
LPR Value
=14.58
X1 = 3
X2 = 2.86

From the above, since the upper and lower bound are equal, the optimal solution to
the problem (original) with only x1 required to be integer has been found. It is given by x 1
= 3 and x2 = 2.86, with an objective function value of 14.58

INVENTORY PROBLEMS
INTRODUCTION
An inventory can be defined as any idle resource of an enterprise. An inventory
problem exists when it is necessary to stock physical goods or commodities for the
purpose of satisfying demand over a specified time horizon (finite or infinite). Almost
every business must stock goods to ensure smooth and efficient running of its operation.
Decisions regarding how much and when to order are typical of every inventory problem.
The required demand maybe satisfied by stocking once for the entire time horizon or by
stocking separately for every time unit of the horizon. The two extreme situations
(overstocking and under-stocking) are costly. Decisions may thus be based on the
minimization of an appropriate cost function that balances the total costs resulting from
over-stocking and under-stocking.
A GENERALIZED INVENTORY MODEL
The ultimate objective of an inventory model is to answer two questions.
1. How much to order?
2. When to order
The answer to the first question is expressed in terms of what we call the order
quantity and the when-to-order decision is the inventory level at which a new order
should be placed usually expressed in terms of re-order point.
The order quantity and re-order pint are normally determined by minimizing the
total inventory cost that can be expressed as a function of these two variables. We can
summarize the total cost of a general inventory model as a function of its principal
components in the following manner:
(Total inventory cost) =

(purchasing cost) + (setup cost) + (holding cost)

+ (shortage cost)

TYPES OF INVENTORY MODELS

In general, inventory models are classified into two categories:
1. Deterministic model and
2. Stochastic model

Definitions:
1. Delivery lags or lead times is the time between the placement of an order and its
receipt and may be deterministic or stochastic.
2. Time Horizon defined the period over which the inventory level will be controlled.
This horizon may be finite or infinite depending on the time period which demand can
be forecast reliably.
3. Stock replacement: Although an inventory system may operate with delivery lags.
The actual replenishment of stock may occur instantaneously or uniformly.
Instantaneous replenishment can occur when the stock is purchased from outside
sources. Uniform replenishment may occur when the product is manufactured locally
within the organization.
DETERMINISTIC MODEL (Single item static model)
The simplest type of inventory model occurs when demand is constant over time
with instantaneous replenishment and no shortages. Typical situations to which this
model may apply are:
1. The use of light bulbs in a building
2. The use of clerical supplies, such as paper, pads and pencil in a large company
3. The consumption of staple food items, such as bread and milk
Fig (1) illustrates the variation of the inventory level. It is assumed that demand occurs at
the rate (per unit time). The highest level of inventory occurs when the inventory level
reaches zero level y/ time units after the order quantity y is received.

y
Inventory
Level

Points in time at which orders are received

Average Inventory y/2

x time

t0 y/

Fig. 1
The smaller the order quantity y, the more frequent will be the placement of new
orders. However, the average level of inventory hold in stock will be reduced. On the
other hand, larger order quantities indicate larger inventory level but less frequent
placement of order (see fig. 2).
Low
Ordering
Frequency

y
Inventory
Level

High
Ordering
Frequency

Fig. 2

x time

Because there are costs associated with placing orders and holding inventory in
stock, the quantity y is selected to allow a compromise between the two types of costs.
This is the basis for formulating the inventory model.
Let K be the set up cost incurred every time an order is placed and assume that the
holding cost per unit inventory per unit time is k. Hence, the total cost per unit time TCU
as a function of y may be written as
TCU (y) = Setup cost/unit time + holding cost/unit time
TCU (y) =

+ h (y/2)

As seen from fig. (1), the length of each inventory cycle is to = y / and the
average inventory in stock is y/2
The optimum value of y is obtained by minimizing TCU (y) with respect to y.
Thus, assuming that y is a continuous variable, we have
= k/y + h/2 = 0
Which yields the optimum order quantity as
y=
(It can be proved that y minimizes TCU(y) by showing that the 2 nd derivates at y is
strictly positive). The order quantity above is usually referred to as Wilsons economic lot
size.
The optimum policy of the model calls for ordering units every to

time units.

The optimum cost TCU (y) obtained by direct substitution is 2k/.

EXAMPLE 1:
The daily demand for a commodity is approximately 100 units. Every time an
order dared, fixed cost is N100 is incurred. The daily holding cost per unit inventory if
N0.02. If the lead time is 12 days. Determine the economic lot size and the re-order point
SOLUTION:
From the earlier formula the economic lot size is
y=

= 2* 100*/0.02 = 1000 units

The associated optimum cycle length is this given as
b0 * =

= 1000/100 = 10 days

Since the lead time is 12 days add the cycle length is 10 days re-ordering occurs when the
level of inventory is sufficient to satisfy the demand for two (- 12, 10) days. Thus the
quantity y* = 1000 is ordered when the level of inventory reaches 2 * 100 = 200 units.
Notice that the effective lead time is taken equal to 2 days rather than 12 days.
This result occurs because the lead time is longer than t 0*.
EXAMPLE 12:
A manufacturer has to supply his customer with 600 units of his product per year.
Shortages are not allowed and the shortage cost amounts to N0.60 per and per year. The
setup cost per run is N80.00. Find the optimum run size and minimum average yearly
cost.
Solution:
Since = 600 units/year
K = N80.00
h = N0.06

y=
2 * 80 * 600 / 0.06 = 400 units opt. run time
And the minimum average yearly cost = 2kh
= 2 * 80 * 60 * 0.60
= N240.00
EXERCISES
1.

XYZ Company purchases a component used in the manufacturing automobile

generators directly from the suppliers. XYZs generator production which is
operated at a constant rate will required 1000 components per month throughout
the year (12,000 units annually). If ordering cost is N25 per order, unit cost is

N200 per component and annual inventory holding costs are charged at 20%.
Answer the following inventory policy question for XYZ
a. What is the economic order quantity (EOQ) for this component?
b. What is the length of cycle time in months?
c. What are the total annual inventory building and ordering cost associated with
your recommended EOQ?
2.

The demand for a particular item is 18,000 units per year. The holding cost per
unit is N1.20 per year, and the cost of the replenishment rate is instantaneous.
Determine
a. Optimum order quantity
b. Number of orders per year
c. Time between orders and
d. Total cost per year when the cost of 1 unit is N1.00

GAME THEORY
A game is defined to be a contest between opponents in which each has a no
(finite of infinite) of courses of action, strategies and the outcome of any combination of
strategies is known beforehand.
Definitions:
1.

A two person zero sum game is one played by 2 persons or groups where the

gain of one person will be exactly equal to the loss of the other so that the sum total of
gains and losses will be equal to zero. A 2 person zero-sum game can be formulated in
the form of a matrix payoff matrix shown below:
Player C Strategies
C1

C2

C3

...

C1

R1

a11

a12

a13

...

a1n

Player R

R2

a21

a22

a23

...

a2n

Strategy

R3

a31

a32

a33

...

A3n

.
.
.
.
...
.
.
.
.
.
.
.
.
.
.
.
Rm
am1
am2
am3
...
amn
The player R controls the rows R1, R2, , Rm which represents his strategies while player
C controls the columns C1, C2, Cn which represents his strategies. If player R chooses
be ith strategy and player C the jth strategy then the element aij is assumed to represent
the payoff from player C to player R i.e. if aij is a +ve no, it represents payment of C to R
and if ve it denotes the payment of R to C.
2.

The element aij of the payoff matrix (of order mx n) of a game is called a saddled

point if it is minimum along the ith row elements and the maximum along the jth column
elements. Thus a game is said to be strictly determined, if an only if it has a saddled
value. The value of the game is equal to the saddle point. The optimal strategies for the 2
players are given by the row that contains the saddled point for the player R, and the
column that contains the saddle point for the Player C.
Determination of Saddle Point Minimax (Maximum) Principle

In any game problem, each player is interest in determining his own optimal
strategy. However, because of the lack of information regarding the specific strategies
selected by the other players, optimality has to be based on a conservative principle so,
each player selects his strategy which guarantees a payoff that can never be worse by the
selection of his opponents, this idea is called the Minimax (or Maximum) principle and
is illustrated below.
Example:
Let the payoff matrix of a 2-person zero-sum game be as in the matrix below. Find
the optimal strategies of the players.
Player C Strategies
C1

C2

C3

C4

Row
Minimum

Player R
strategy

R1

R2

R3

-3

-3

Column

Maximum
If player R selects his strategy R1, he may gain 7, 1, 8, 4 depending on the strategy
selected by player C. However, player R is guaranteed a gain of at least 1 = (7, 1, 8, 4)
irrespective of the strategy of C. Similarly R is guaranteed a gain of at least 42min (5, 4,
6 7) if he choose strategy R2 and at least -3 = min (6, 2, -3, 6) if strategy R3 is selected.
Thus if player R wants to maximize his gain irrespective of strategy selected by C, he has
to maximize the minimum gain i.e. max (1, 4, -3) = 4. Thus the strategy R2 is to be
chosen by R based on the maximum principle with 4 as the maximum value of the game.
On the othe hand, if player C chooses strategy C, he loses 7, 5, 6 depending on the
strategy selected by player R. however he can lose no more than 7 = max (7, 5, 6)
regardless of Rs strategies. In similar manner, player C loses no more than 4 = max (1, 4,
2), 8 = max (8, 6, -3) and 7 = max (4, 7, 6) by choosing strategies C 2, C3 and C4
respectively regardless of the strategies selected by R. Thus C selects the particular

strategy which minimizes his maximum losses irrespective of the strategies of R, i.e. min
(7, 4, 8, 7). The minimum of maximum loss is given by strategy C 2 based on minimax
principle.
A game as in the e.g. above, where the minimax value = maximin value the
corresponding pure strategies are called optimal strategies and the game is said to have a
saddle point.
As illustrated below, it is possible to have a situation where there is no saddle
point and hence one cannot find a pure strategy solution consider a game for which the
pay off matrix is as below:
Player C Strategies

Player R
strategy

C1

C2

C3

Row Minimum

R1

R2

R3

Column

Maximum
So, in general maximin value of game minimax value. (3. 27)
As in the last example above, mixed strategies in which both players will search
for a correct strategy mixture to find equilibrium, has to be used. The value of the game
at equilibrium is uniquely determined by the right strategy mixture and satisfies the
inequality (3.27).
The correct strategy mixture for each player is determined by assigning to each
strategy its probability of being chosen. Let r1, r2, .. rn represents the probabilities
with which player C selects the pure strategies R1, R2 . Rm respectively and let
C1, C2 . Cm be the probabilities with which player C selects the pure strategies
respectively. The sum of the probabilities for the strategies of each player may be equal
to one i.e.
m

 ri = 1,

ri 0 i = 1, 2, , m

i=1

and

(3.28)

ci = 1,

ci 0 j = 1, 2, , n

j =1

The solution of mixed strategy problems is also based on the minimax principle in
the sense that R selects the values of ri so as to maximize the smallest expected value of
pay off in a column, where as C chooses the values of c j so as to minimize the largest
expected value of pay off in a row.
Thus, the player R finds ri which will
m

maximize

minimize of

ai1 ri, ai1 ri, ain ri,

i=1

i=1

i=1

w.r.t ri with r, 0
m

and ri = 1

(3.29)

i=1

and the player C selects cj which will

n

maximize

minimize of

aij cj, a2j cj, amj cj,

i=1

i=1

i=1

w.r.t ci with r, 0
m

and ri = 1

(3.30)

j=1

Let ri (i = 1, 2, , m) and c j (j = 1, 2, , n) denote the optimal solution, since each pay

off element aij is associated with a probability combination (r i, cj) the optimal expected
value of the game is given optimum expected value of the game = m n
and

= aijr ic

(3.31)

i=1 j=1

Solution of a Game Problem

Equation (3.29) can be transformed to a LPP as follows
m

Let v = min ai1 ri, ain ri,

i=1

i=1

Then (3.29) becomes

Maximize f (r1, r2, rm) = v subject to
m

ai1 ri, v
i =1
m

ai2 ri, v

(3.32)

i=1

.
.
.
m

ain ri, v
i=1

And ri 0, i = 1, 2, m
Clearly, (3.32) is a LPP, whose solution gives the value of the game. The simplex method
is used to solve the game problem below
Example
There are 2 competing departmental stores R and C in a city. Both the stores
customers are equally divided between the two. Both the store, plan to run annual
discount sales in the last week of December for this they want to attract more number of
customers by using advertisement through newspaper, radio and television. By seeing the
market trend, the store R constructed the following pay off matrix below, where the
number in the matrix includes a gain or loss of customer.

Strategies of Store C

C1

C2

C3

Row Minimum

Strategies
Store C

R1

40

50

-70

-70

R2

10

25

-10

-10

R3

100

30

60

30

Column

100

50

60

Maximum

Assuming that a gain of customers to store R means a loss to C, find the optimal
strategies for both the store along with the value of the gains.
Solution
The minimax and maximin values for this pay off matrix are 50 and 30 respectively and
hence the game does not have saddle point. Now, dividing the constraint in (3.32) by v
we have
Maximize f (r1, r2, rm) = v
m

ai1 ri, 1
i =1

ai2 ri, 1
i=1

(3.33)

.
.
.
m

ain ri, v
i=1

And ri 0, i = 1, 2, m
Defining new variables xi as
xi = ri i = 1, 2, m
v
Note that
Maximum of v = minimum of 1 = minimum of
v

xi
i=1

The LPP (3.32) can be restated as

Maximize f(x1, x2, xm) = v

xi subject to
i=1
m

ai1 xi, 1
i =1
m

ai2 xi, 1
i=1

(3.33)

.
.
.
m

ain xi, v
i=1

And xi 0, i = 1, 2, m
Respectively from (3.30) we have the LPP
Maximize g(y1, y2, ym) =

yj

subject to

j =1

a1j yj, 1
i =1
m

a2j yj, 1
i=1

(3.34)

.
.
.
m

a1j yj, 1
i=1

And yi 0, j = 1, 2, n
Where g = 1 , yj = cj , j = 1, 2, n
v
v
For this example, the problem of store C can state as LPP as follows:
Maximize

g = y1 + y2 + y3 subject to

a11y1 + a12y2 + a12y3 1

a21y1 + a22y2 + a23y3 1
a31y1 + a32y2 + a33y3 1
y1 0, i = 1, 2, 3

And

This is restated as
Maximize

f = y1 + y2 + y3 subject to
40y1 + 50y2 70y3 + y4 = 1
10y1 + 25y2 10y3 + y5 = 1
100y1 + 30y2 + 60y3 + y6 = 1
y1 0, i = 1, 2, 3

And
Basis

CB

y2
+1
50

y3
+1
-70

y4
0
1

y5
0
0

y6
0

y1
+1
40

y4

y5

10

25

-10

y6

100

30

60

zj

cj - zj

+1

+1

+1

y1
+1

y2
+1
38

y3
+1

y4
0

y5
0

-94

-16

y6
0
4
10
1
10

Basis

CB

y4

y5

0
0

22

Ratios

1
100

1
40
1
10
smallest value

Ratios
6
10
9
10

6
smallest value
380
9
220

y1

1
zj

3
10
3
10

3
5
3
5

7
10

y1
+1

y2
+1

1
0

cj - zj

Basis

CB

y2

y5

y1

1
zj
cj - zj

Basis

CB

y4

y5

y6

1
zj
cj - zj

y1
+1
94
51
-1460
51
38
51
132
51
-81
51

2
5

-1
100

y3
+1
-47
19
730
19
51
38
43
38
81
38

y4
0
1
38
-11
19
-3
360
7
380
-7
380

y5
0

y6
0
-1
95
5
38
1
76
1
380
-1
380

y2
+1

y3
+1

v* = 1 = 34
g
Thus, the optimum solution for C
v* = - 1 = 34,
g

1
100
1
100

y4
0
1
85
-114
323
-1
170
1
170
-1
170

0
1
0
0
0

y5
0
0
1
0
0
0

y6
0
7
510
-25
102
1
102
2
85
-12
510

1
100
1
100

1
30

Ratios
3
190
21
21
38
1460
1
1
190 255 smallest value
2
95

Ratios
13
510
41
102
1
255
1
34

c2* = y2* v* = 13 * 34 = 13
510
15
c3* = y3* v* = 1 * 34 = 2
510
15
c1 * = 0
The optimal strategy for R is given by
v* = 34,
r1* = x1 *v* = 1 *34 = 1
170
5,
r2* = x2 *v* = 0,
r3* = 2 *34 = 4
85
5
Thus, store R will gain 34 customers from C when both apply the optimal advertising
strategies during the annual reduction sale period.

Exercises
1.

Find the optimal strategies and the value of the 2 games whose payoff matrices are
given below:
C1

C2

C3

C4

R1

-12

R2

-3

R3

-5

-2

10

-3

13

12

R4

-2

-5

P1/P2

T1

T2

T3

T4

S1

12

14

S2

-10

S3

-4

S4

2.

An m x n matrix is called a Latin square if each row and column contains each of
the integers from 1 to m. Show that a game which has this as its payoff matrix has
the value (1 + m)

3.

Two players fight a duel, they face other 2n paces apart and each has a single
bullet in his gun. At a signal each may-fire, if either is hit or if both fire the game
ends. Otherwise both advance one pace and may again fire. The game of course

ends anyway by the time n paces has been taken. The probability of either hitting
his target if he fires after the ith pace forward is 1/n. The payoff is +1 to a player
who survives after his opponent is hit, and 0 if neither or both are hit; the guns are
silent so that neither knows whether or not his opponent has fired.
Show that if n = 4, the strategy shoot after taking two steps is optimal for both but
that if n = 4, a mixed strategy is optimal.

DECISION THEORY
The ultimate purpose of any operations research analysis is to enable operations to
be run efficiently and effectively, and this in turn involves selecting the best of the
alternative ways and means of conducting operations. Fundamental, then, to any
operations research exercise is the final step of making a decision between alternatives,
and the principles underlying such decision making are referred to as Decision Theory
Structuring the Decision Problem
To illustrate the decision analysis approach, let us consider the case of Political
Systems, Inc (PSI), a newly formed computer service from specializing in information
services such as surveys and data analysis for individuals running for political office. The
firm is in the final stages of selecting a computer system for its Midwest branch, located
in Lagos. While PSI has decided on a computer manufacturer, it is currently attempting to
determine the size of the computer system that would be most economical. We will use
decision theory to help PSI make its computer decision.
The first step is to identify the alternatives considered by the decision maker. For
PSI, the final decision will be to select one of the three computer systems, which differ in
size and capacity. The three decision alternatives denoted by D 1, D2, and D3 are as
follows:
D1 large computer system
D2 Medium computer system
D3 Small computer system
The second step is to identify the future events that might occur. There events,
which are not under the control of the decision maker, are referred to as the States of
nature. Thus, the PSI states of nature denoted S1 and S2 are as follows:
S1 high customer acceptance of PSI services
S2 low customer acceptance of PSI services
Given the three decision alternatives and the two states of nature, which
computer system should PSI select? To answer this question, we will need information

on the profit associated with each combination of a decision alternative and a state of
nature.
Payoff Tables
We denote the decision alternatives by D1, D2, , Dm, the states of nature by S1,
S2, Sn; and the return associated with decision Di and state Sj by Vij (I = 1, 2, , m i
j = 1, 2, , n). A process requiring the implementation of just one decision is defined
completely by Table 1. A table of this form is referred as a payoff table. In general,
entries in (a) can be stated in terms of profits, costs etc. Using the best information
available, management has estimated the payoffs or profits for the PSI problem. There
estimates are presented in Table 2
Table 1: States of Nature
S1

S2

Sn

D1

V11

V12

V1n

D2

V21

V22

V2n

Dm

Vm1

Vm2

Vmn

Table 2:
Decision alternatives

High Acceptance S1

Low Acceptance S2

Large system

D1

200,000

20,000

Medium system

D2

150,000

20,000

Small system

D3

100,000

60,000

Decision Trees
A decision tree provides a graphical presentation of the decision-making process.
Figure 1 shows a decision tree for the PSI problem. Note that the three shows the natural
or logical progress that will occur overtime.

Figure 1
High (S2)

200,000
Large (D1)
2
Low (S2)
-20,000
High (S1)
150,000
Medium (D2)

3
Low (S2)
20,000
High (S1)
100,000
Small (D3)
4
Low (S2)

60,000
Using the general terminology associated with decision trees, we will refer to the
intersection or junction points of the tree as nodes and the arcs or connectors between the
nodes as branches. Fig 1 shows the PSI decision tree with the nodes numbered 1 to 4.
When the branches leaving a given node are decision branches, we refer to the nodes as
decision node. Decision nodes are denoted by squares. Similarly, when the branches
leaving a given node are state-of-nature branches, we refer to the node as a state-ofnature node. State-of-nature nodes are denoted by circles. Using the node-labelling
procedure, node 1 is a decision node, where as nodes 2, 3 and 4 are states of nature nodes.
Decision Making without Probabilities
This section consider approaches to decision making that do not require
knowledge of the probabilities of the states of nature
Optimistic Approach
The (~) evaluates each decision alternative in terms of the best payoff that can occur. The
decision alternative that is recommended is the one that provides the best possible payoff.
For a problem in which maximum profit is desired, as it is in the PSI problem, the
optimistic approach would lead the decision maker to choose the alternative
corresponding to the largest profit. For problems involving minimization, this approach

leads to choosing the alternative with the smallest payoff. To illustrate the use of the ~,
we will show how it can be used to develop a recommendation for the PSI problem.
Table 3
Decision alternatives

Maximum Payoff

Large system

D1

200,000

Medium system

D2

150,000

Small system

D3

100,000

Maximum of the maximum

payoff values

Conservative Approach
The conservative approach evaluates each decision alternatives in terms of the
most payoff that can occur. The decision alternative recommended is the one that
provides the best of the worst possible payoffs. For a problem in which the output
measure is profit, as it is in PSI problems. The (~) would lead the decision maker to
chose the alternative that maximizes the minimum possible profit that could be obtained.
For problems involving minimization, this approach identifies the alternative that will
minimize the maximum payoff.
Table 4
Decision alternatives

Maximum Payoff

Large system

D1

-20,000

Medium system

D2

20,000

Small system

D3

60,000

Maximum of the maximum

payoff values

Minimax Regret Approach

(~) is another approach to decision making with certainty. This approach is neither
purely optimistic nor purely conservative. We illustrate the (~) for the PSI problem. In
maximization problem, the general expression for opportunity loss or regret is given by
the formula:
Opportunity loss or Regret
Rij = Vj Vij
Where

---

(*)

Rij = reject associated with decision alternative Di and state of nature Sj

Vj = best payoff value under state of nature Sj
Vij = payoff associated with decision alternative D1 and state of nature Sj
Using eq (*) and the payoff in Table 2, we can compute the regret associated with all
combinations of decision alternatives Di and States of nature Sj
Table 5
Regret or opportunity loss for
the PSI problem

States of Nature

Decision Alternatives

High Acceptance
S1

Low Acceptance
S2

Large System

D1

80,000

Medium System

D2

50,000

40,000

Small System

D3

100,000

Table 6
Decision Alternatives

Maximum Regret or Opportunity Loss

Large system

D1

80,000

Medium system

D2

50,000 Minimum of the maximum

Small system

D3

100,000

regret

For the PSI problem, the decision to select a medium-computer system, with a
corresponding regret of N50,000, is the recommended minimax regret decision.
Rank: In cost minimization problems, Vj will be the smallest entry in column j, and
equation (*) must be changed to
Rij = Vij Vj

Decision Making with Probabilities

In many decision-making situations, it is possible to obtain probability estimates
for each of the states of nature. When such probabilities are available, the expected value

approach can be used to identify the best decision alternative. The expect value approach
evaluates each decision alternative in terms of its expected value. The recommended
decision alternative is the one that provides the best expected value.
Let
N = the number of states of nature
P (Sj) = the probability of state of nature Sj
P (Sj) 0 for all states of nature
P (Sj) = P (S1) + P (S2) + + P (SN) = 1
Expected Value of Decision Alternative Di

Using the payoff values Vij shown in Tale 1 and supposes that PSI management believes
that S1, the high acceptance state of nature, has a 0.3 probability of occurrence and that
S2, the low-acceptance state of nature, has a 0.7 probability. Thus, P (S1) = 0.3 and P (S2)
= 0.7 and equation (**), expected values for the three decision alternatives can be
calculated:
EV (D1) = 0.3 (200,000) + 0.7 (-20,000) = N46,000
EV (D2) = 0.3 (150,000) + 0.7 (20,000) = N59,000
EV (D3) = 0.3 (100,000) + 0.7 (60,000) = N72,000
Thus, according to the expected value approach, D3 is the recommended decision
since D3 has the highest expected value (N72,000)

Using Decision Tree

Figure 2
High (S2)
200,000
P (S1) = 0.3
2
Low (S2)
-20,000
P (S2) = 0.7
High (S1)
150,000
P (S1) = 0.3
1

3
Low (S2)
20,000
P (S2) = 0.7
High (S1)
100,000
P (S1) = 0.3
4
Low (S2)
60,000
P (S2) = 0.7

PSI Decision Tree with State-of-Nature Branch Probabilities

Large (D1)
2

EV = 0.3 (200,000) + 0.7 (-20,000) = N46,000

EV = 0.3 (150,000) + 0.7 (20,000) = N59,000

EV = 0.3 (100,000) + 0.7 (60,000)= N72,000

Medium (D2)

Applying the Expected Value Approach Using Decision Trees

Sensitivity Analysis
In this section, we consider how changes in the probability estimates for the states
of nature affect or utter the recommended decision. The study of the effect of such
changes is referred to as ~. One approach to ~ is to consider different probabilities for the
states of nature and then recompute the expected value for each decision alternative.
Repeating this several times, we can begin to learn how changes in the probabilities for
the states of nature affect the recommended decision. For example, suppose that we
consider a change in the probabilities for the states of nature such that P (S1) = 0.6 and P
(S2) = 0.4. Using these probabilities and repeating the expected value computations, we
find the following:
EV (D1) = 0.6 (200,000) + 0.4 (-20,000) = N112,000
EV (D2) = 0.6 (150,000) + 0.4 (20,000) = N98,000
EV (D3) = 0.6 (100,000) + 0.4 (60,000) = N84,000
Thus, with these probabilities, the recommended decision alternative is D 1, with an
expected value of N112,000.
The only drawback to this approach is the numerous calculations required to
evaluate the effect of several possible changes in the state-of-nature probabilities.

DYNAMIC PROGRAMMING AND MULTISTAGE OPTIMIZATION

Optimization problems consist in selecting from among the feasible alternatives
one which is economically optimal. A problem of this nature is solved by formulating a
mathematical model of the problem, typically a maximization model in which a
preference function is to be maximized subject to a number of side conditions, and
applying a method of solution tailored to the particular kind of problem. The variables of
the model, interdependent through the side relations, are determined simultaneously in
the solution.
Consider, for example, the linear programming problem
f=

8x1 + 10x2 = max

4x1 + 2x2 12

(1)

x1 , x2 0
to be interpreted as a problem of optimal capacity utilization. x 1 and x2 are quantities
produced per period of two commodities which require 4 and 2 machine hours per unit,
and the right-hand side of the side condition is maximum available machine time per
period, the coefficients in the preference function are unit profits. Solving by the simples
method (or, what is simpler in such a trivial case, geometrically or by numerical
inspection) we get.
x1 =0, x2 =, f = 60
The optimal value of the two decision variables is found simultatiously in the
solution procedure.
An alternative approach is to determine the variable one at a time (sequentially),
decomposing the problem into a series of stages each corresponding to a sub problem in
only one variable, and solving the two single-variable problem (1). This is the basic idea
underlying dynamic programming (DP).

The decomposition of the problem (1) can be illustrated as shown in fig. 8. Let us
assume for convenience that commodity no. 1 is produced first (stage 1). We might as
well have started with the second commodity; the order in which they are arranged is
purely formal in a case like this where the decomposition into stages does not reflect a
sequence in time.
Now, for the production of the first commodity, 12 units of the capacity
factor (machine hours) are available as shown in the flow diagram. If x1 units are
produced, 12 4x1 machine hours are available as input for the second stage. After
producing x2 units of commodity
x1

12

Stage 1

x2
12 4x1

Stage 2

8x1

12 4x1 2x2

10x2

Figure 8

No 2, we are left with 12 4x1 2x2 machine hours, corresponding to the slack variable
in (1) which represents unutilized capacity. The two stages contribute 8x 1 and 10x2
respectively to total profit.
We can now solve the problem backwards, treating x1 as a parameter and
optimizing stage 2 with respect to the variable x2. For parametric x1 = x2 the maximal
capacity left to stage 2 is 12 4x1 machine hours so that the (parametric) subproblem of
stage 2 is

f1 = 10x2 = max
2x2 12 4x1
x2 0

(3)

which is a linear programming problem like the total problem (1), only it is singlevariable problem. The solution is obviously.
x2 = 6 2x1; f1max = 60 20x1

(4)

where x1 is a parameter.
Next we optimize stage 1 with respect to its decision variable, x1. The capacity
available is 12 machine hours. Production of x1 units contributes 8x1 to total profit, but
against this we have to consider that machine hours left over and used by stage 2 also
affect total profit, contributing f1max = 60 20x1 which also depends on x1. The
optimization problem of stage 1 becomes
f2 = 8x1 + f1max = 60 12x1 = max
4x1 12

(5)

x1 0
which is also a linear problem. Because f2 which expresses profit contributed by the
first stage plus the (parametric) maximum profit earned by the second is a decreasing
function of x1, the solution obviously is
x1 = 0; f2max = 60

(6)

Having thus found the optimal value of x1, which is also a parameter in the solution for
x2, we substitute it into (4) to get
x2 = 6 2x1 = 60

(7)

The solutions which we have found for the two variables are seen to agree with (2), and
total profit f = 60 is seen to be equal to f2max as it should be seen f2max was calculated as
the total of stage contributions to profit. What we did in solving (5) was to maximize the
profit of stage 1 a function of x1 plus the maximum profit earned by stage 2 for any
given value of x1.
In this way we have solved an optimization problem in two variables by
transforming it into a series, or sequence, or two single-variable problems. This is an
example of dynamic programming. The subproblems corresponding to the individual
stage are of the same type as the total problem (1) in the present case, a linear
programming problem and they are solved by the same method as that applied in the

simultaneous solving of the total problem (e.g. the simplex method). In other words,
dynamic programming does not refer to a particular class of optimization problems
(e.g. linear programming problems) or to a specific method of solution (like the simplex
method); rather, it indicates a general procedure for decomposing a problem into a series
of subproblems involving fewer variables, and combining their solutions to get the
solution of the original problem.
3. When an optimization problem is formulated as a multistage problem to be
solved by dynamic programming, it is convenient to introduce state variables yn
associated with the individual stages (numbered n = 1, 2, .., n). The production process of
Figure 8 may be thought of as starting in a (given) initial state where y0 = 12 machine
hours are available; this is the input state of stage 1. Producing x1 units of the first
commodity, each of which requires 4 machine hours, changes the state of the system:
available capacity is reduced by 4x1 machine hours so that the output state of stage 1
which is also the input state to stage 2 becomes y1 = y0 4x1 = 12 4x1. Producing x2
units of the second product, available capacity is further reduced to y 2 = y1 2x2 = y0
4x1 2x2, which in this case represents the final state. (y2 0 by definition but otherwise
unknown).
Thus, the input state of stage no. n, yn-1, is transformed into an output state yn, the
change being brought about by the decision variable
x1

y0 = 12

Stage 1

x2
y1 = y0 - 4x1

Stage 2

y2 = y1 2x2

f2 = 10x2

f1 = 8x1

Figure 9
of the stage, xn. The successive changes of the state of the system can formally be
described by transformation equations of the form
yn = tn (yn-1, xn) (n = 1, 2, , n).

In the example they have the form

y1 = y0 4x1
y2 = y1 2x2

(8)

together with the nonnegativity requirement x 1, x2, y1, y2 0 they are equivalent to the
restrictions of the original problem (1), y0 being = 12 and y2 representing the slack
variable.
The stage returns, i.e. the contributions of the individual stages to the preference
function f, will in the general case depend on the input state and the decision variable:
rn = rn (yn 1, xn) (n = 1, 2, , n);
in the present example these return functions are of the simple form
r1 = 8x1
r2 = 10x2

(9)

Introducing these symbols into Fig. 8, the flow diagram of the two-stage problem
has the form of Fig. 9. The backward solution now proceeds as follows.
At the first stage of the calculations corresponding to the last stage in the
production system, n = N = 2 the input style y1 is considered as a parameter, inherited
from the previous stage of the system. The stage is optimized by maximizing its decision
function, f1 here equal to the stage return, r2 (x2) subject to the parametric capacity
restriction 2x2 y11:
f1 = r2 (x2) = 10x2 = max
2x2 y1
x2 0

(10)

The solution to this parametric single-variable LP problem is

x2 (y1) = 0.5y1
f1 (y1) = 5y1

(11)

where f1 denotes the maximum value of the stage decision function, F 1 = f1max
At the second stage of the computations (production stage 1) we maximize the
decision function f2 = r1 (x1) + f1 (y1) subject to the capacity limitation 4x1 y0 (i.e., y1 =

y0 4x1 0); substituting the stage transformation y1 = y0 4x1, f2 becomes a function of

x1 and y0 so that we have the LP problem.
f2 =

r1 (r1) + F1 (y1) = 8x1 + 5y1

8x1 + 5 (y0 4x1) = 5y0 12x1 = max

(12)

4x1 y0
x1 0.
The solution is
x1 (y0) = 0
f2 (y0) = 5y0

(13)

when F2 = f2max.
The solution to the complete problem (1) called the optimal policy can now be
determined by solving the recursive equation system formed by the parametric optimum
solutions (11) and (13) and the transformation equations (8), starting from the initial state
y0 = 12:
Transformation
Equations
yn = tn (yn-1, xn)

Parametric Optimal
solutions
xn = xn (yn-1)

Maximum of decision
function

y1 = y0 4x1 = 12

x1 (y0) = 0

f2 (y0) = 5y0 = 60

y2 = y1 2x2 = 0

x2 (y1) = 0.5y1 = 6

f1 (y1) = 5y1 = 60

y0 = 12

where the direction of the computations is the opposite of that followed above in the
optimization of stages. The optimal policy emerges as x1 = 0, x2 = 6 and total profit is f =
F2 = 60. If y0 had been a parameter, the solution now parametric would have been x1 =
0, x2 = 0.5y0, f = 5y01.
The decomposition by which we solved problem (1) can be described as follows.
Replacing the side conditions by the equivalent formulation
y1 = y0 4x1, y1 0
y2 = y1 2x1, y2 0,
this together with the nonnegativity requirements implies

0 x1 y0

(14)

0 x2 y1

(15)

Then we can write (1) in the form

Fmax = max (8x1 + 10x2) = max (8x1 + max 10x2)
x1, x2

x1

x2

subject to (14) (15); clearly this maximization problem can be decomposed into two
single-value problems, corresponding to (10) and (12):
F1 (y1) = max 10x2 (0 x2 y1)
F2 (y0) = max {8x1 + F1 (y1)}
= max {8x1 + F1 (y0 4x1)} (0 x1 y0),
Where fmax = F2 (y0). In a more general formulation the decomposition of a two-stage
problem
Fmax = max [r1 (y0, x1) + r2 (y1, x2)]
can be expressed in the recursion equations
F1 (y1) = max [r2 (y1, x2)]

(16)

F2 (y0) = max [r1 (y0, x1) + F1 {t1 (y0, x1)}].

(17)

This procedure of solving a dynamic programming problem by backward

recursion can be generalized to any number of variables. A flow diagram for an N-stage
system is shown in Fig. 10.

x2

x1
y0

y1 = t1 (y0, x1)

f1 = (y0, x1)

xn
y2

f2 = (y1, x2)

yn-1

xN
yn

yN-1

fn = (yn-1, xn)

yN

fN = (yN-1, xN)

Figure 10
The decision functions of stages N, N 1, , 2, 1 are respectively
f1 = rN (yN 1, xN)
f2 = rN -1 (yN-2, xN-1) + F1 (yN-1) where yN-1 = tN-1 (yN-2, xN-1)

(18)


fN = r1 (y0, x1) + FN-1 (y1)

where y1 = t1 (y0, x1),

Fj being the maximum of f j (j = 1, 2, , N). Maximizing the decision function of each

stage with respect to its decision variable, treating the input state as a parameter, we get
the parametric stage solutions
xn = xn (yn 1) (n = N, N 1, , 1)

(19)

which can be sewn together by means of the transformation equations so that we get
the parameters determined. y 0 now determines x1 and together they determine y1; this
gives x2 which with y1 determines y2; and so forth as illustrated by Fig. 11
y0

y1 = t1 (y0, x1)
x1 (y0)

y2 = t2 (y1, x2)
x2 (y1)

x3 (y2)

Figure 11
It follows from (18) that the maximum of fN represents the accumulated value of
the stage returns, e.g. FN = fmax.
If the variables of a DP problem are allowed to take discrete values only, or if the
stage returns and/or the transformation functions are given in tabular form for discrete
values of the variables, the problem will have to be solved by tabular computations.
For example, the decision variables may be required to have integral values (xn =
0, 1, 2, ) because the interpretation of the problem is such that fractional values would
be meaningless. This is so in problem (1): strictly speaking it is impossible to produce a
fractional number of units of a commodity, e.g. x1 = 2.6. Thus, although in this case the
analytical shape of the return functions is known, r1 and r2 are defined only for integral
values of x1 ( 3 ) and x2 ( 6):

x2

x1

r1 (= 8x1)

16

24

r2 (= 10x2)

10

20

30

40

50

60

The transformation functions in tabular form are as follows:

____________________________________
y1 (= y0 4x1)
y0\x1

12

12

______________________________________________________
y2 (= y1 2x2)
y1 \ x2

12

12

10

where y1 is confined to the values 0, 4, 8, and 12 resulting from the first table. The black
cells in the last table correspond to combinations of values of y 1 and x2 which are not
feasible because they would imply a negative value of y2 (cf. the sign restriction y2 = y1
2x2 0).
The solution procedure, using backward recursion, now proceeds as follows. For
the last stage (n = N = 2) we have:
Stage 2

f1 = r2 (x2) (= 10x2)

y1 \ x2

0 1 2 3 4 5 6

F1 (y1)

x2 (y1)

y2 (y1)

0 10 20

20

0 10 20 30 40

40

12

0 10 20 30 40 50 60

60

where the maximal value of the decision function f1 for each of the possible input states
(values of y1) is shown in bold-faced type. This parametric optima and the optimal values
of the decision variable are listed in the F1 and x2 columns to the right. The last column

gives the resulting values of the output state y2, computed from the transformation
function (or table).
For stage 1, we have the decision function f2 = r1 (x1) + F1 (y1) where the
transformation gives y1 for each value of x1; for example, x1 = 2 implies y1 = y0 4x1 = 4,
and the preceding table then gives F1 (y1) = 20. The stage computations are done in the
following table:
Stage 1

f2 = r1 (x1) + F1 (y1)
{ = 8x1 + f1 (y1)}

y0 \ x1

12

0+60 8+40 16+20 24+0

F2 (y0)

x1 (y0)

y1 (y0)

60

12

These tables correspond to (10) (11) and (12) (13) respectively, and the
optimal solution can be found similarly, starting with the last table. y0 = 12 gives x1 = 0
(the optimal stage solution) which leads to y1 = 12 (by transformation equation).
Proceeding to the first table, y1 = 12 (as just determined) gives the stage optimum x2 = 6
and the output state y2 = 0. These values are indicated in italics. The maximal total return
is f2 (y0) = 60.
It will now be clear why it was expedient to solve the decomposed version of
problem (1) backwards, starting with the optimization of the last stage. The procedure led
to a recursive system which has the initial state y0 as it starting point as shown in Fig.
11 and it was y0 that was given (y0 = 12). This suggests that problems in which the final
state yN is given may be solved in the opposite direction, proceeding forwards from the
first stage.
To show how this is done, let us redefine the state variables y n in the example so
that the final state y2 now represents total accumulated use of capacity, including idle
capacity; the latter is put first as the input state y 0 of the first stage of the production
system. Then y1 represents the accumulated use of capacity, including capacity not
utilized, after the first commodity has been produced. This leads to the transformations.
y1 = y0 4x1
y2 = y1 2x2

(20)

where y0 0 and y2 = 12. The stage return functions are the same as above, (9).
This dynamic programming problem is another decomposed version of problem
(1). To solve it by forward recursion we reverse the direction
x1

y0 = y1 - 4x1

x2
y1 = y2 - 2x2

Stage 1

Stage 2

y2 = 12

r2 = 10x2

f1 = 8x1

Figure 12
of the system as shown in Fig. 12, where yn is now to be formally treated as the input
state of stage n (n = 1, 2) whereas the yn-1 become output states; we therefore write the
transformation functions (20) in the inverse form
y0 = y1 4x1
y1 = y2 2x2

(21)

(more generally,
yn-1 = tn* (yn, xn)
where tn* is the inverse transformation equation of stage n).
the procedure starts with the optimization of stage 1. The decision function f 1 = r1
(x1) is to be maximized subject to y0 = y1 4x1 0, x1 0, where y1 now the input state
of the stage is a parameter:
f1 = r1 (x1) = 8x1 = max
4x1 y1

(22)

x1 0;
the solution depends on the parameter,
x1 (y1) = 0.25y1

(23)

F1 (y1) = 2y1
where F1 = f1max. At the second stage we have the decision function f2 = r2 (x2) + F1 (y1)
where y1 = y2 2x2 0, x2 0 so that the stage optimization problem becomes

f2 = 10x2 + 2y1 = 6x2 + 2y2 = max

(24)

2x2 y2
x2 0.
The solution is
x2 (y2) = 0.5y2
F2 (y2) = 5y2

(25)

Working backwards from y2 (= 12) through the recursive system (21), (23), (25) we
obtain the optimal solution:
y2 = 12
y1 = y2 2x2 = 0

x2 (y2) = 0.5y2 = 6

y0 = y1 4x1 = 0,

x1 (y1) = 0.25y1 = 0,

and fmax = F2 (y2) = 5y2 = 60.

In the general case of N stages the decision functions are
f1 = r1 (y0, x1)

where y0 = t1* (y1, x1)

f2 = r2 (y1, x2) + F1 (y1)

where y1 = t2* (y2, x2)

fN = rN (yN-1, xN) + FN-1 (yN-1) where yN-1 = tN* (yN, xN).

The parametric stage solutions
xn = xn (yn)

(n = 1, 2, , N)

and the inverse transformation equations determine the optimal policy for given y N as
shown in Fig. 13
It is often possible to solve by forward recursion when the initial state y 0 is given,
or to apply backward recursion to DP problems with given final state y N, using a slightly
modified procedure. However, when the order of the stages is arbitrary and the
transformations can be inverted so that we are free to choose the direction, backward
recursion is generally a more efficient procedure for given y0, and forward recursion for
given yN, (which amounts to the same thing, the only difference being the numbering of
stages and variables).

yN

yN-1 = tN* (yN, xN)

xN (yN)

.
xN-1 (yN-1)

Figure 13
In order for an optimization problem to be solvable by dynamic programming, the
technical structure of the problem (as represented by the restrictions) must be such that
it can be described by a series of successive changes of the state of the system, from the
initial state y0 to the final state yN, each change being effected by a particular decision
variable. The two-stage system of Fig. 9 above is an example of this; the transformation
equations (8) together with x1, y1, x2, y2 0 and y0 = 12 are an equivalent reformulation
of the restriction in problem (1), i.e., 4x1 + 2x2 12 and x1, x2 0, so that the set of
feasible solutions is preserved. In some cases the multistage structure represents a
sequence in time hence the name dynamic programming where the stages
correspond to actual processes and the direction indicates the order in which the
transformations take place. This would be so in our example (1) if the two commodities
were produced in separate processes and commodity no. 1 had to be made first. In many
applications, however, the sequence of stages and the order in which the system passes
through them are an artificial device, introduced in order make the problem solvable by
DP methods. In either case we can choose between forward and backward recursion if the
direction is mathematically arbitrary.
Decomposition of a problem also requires that the objective function satisfies
certain conditions. In general, the function to be maximized in some function of the stage
returns rn,
f = {r1 (y0, x1), , rN (yN-1, xN)}.

(26)

It can be shown that two conditions on the function separability and

monotonicity together are sufficient for decomposition, i.e., for solution by means of a
system of recursive equations. These conditions are automatically satisfied by a class of
functions including the case of additive returns,
f = r1 (y0, x1) + + rN (yN-1, xN).

In this case, as we have seen above, the objective function is obviously

decomposable and the recursion equations have the form (16) (17) for a two-stage
problem, readily generalized to any number of stages [cf. (18)].
The recursion equations can be thought of as a mathematical expression of an
intuitive principle known as the principle of optimality. Consider the two-stage
problem shown in Fig. 9, and let (x1, x2) = (x1, x2) be the optimal policy. The first
decision, x1 = x1, changes the state of the system from the initial state y0 = 12 to y1 = y0
4x1 = y1. The principle of optimally now says that the remaining decision, x 2 = x2, must
represent an optimal policy with respect to the state y1, i.e. it must be an optimal solution
to the remaining one-step DP problem with the initial state y1. The proof is simple: if this
were not so, (x1, x2) = (x1, x2) could not be an optimal policy.
The backward recursive procedure, as expressed in (16) (17), by which we
solved the problem, is based directly on this principle. Starting by optimizing the last
stage (n = 2), we do not know its input stage y1, but we do know that whatever it is i.e.,
whatever the first decision is r2 must be optimal with regard to y1. Hence we optimize f2
= r2 (y1, x2) for parametric y1 as expressed in (16). Proceeding backwards to stage one, we
optimize f2 = r1 (y0, x1) + F1 (y1), i.e. the return of stage 1 plus the parametric optimal
return of stage 2, where y1 = t1 (y0, x1).
The principle of optimality in its general form states that any part of an optimal
policy must be optimal; specifically, the decisions remaining after stage no. n (i.e., x n+1,
, xN) constitute an optimal policy for the series of stages n + 1, , N with regard to the
state yn resulting from the first n decisions.
The multistage structures dealt with above (cf. Fig. 10) are serial systems, i.e.,
ordered sequences of stages where the output state of stage no. n is the input state to stage
no. n + 1. Moreover, they are special in that there is only one decision variable xn and
only one (output) state variable yn for each stage.
As an obvious generalization, the xn and the yn may be vectors so that there are
several decision and state variables per stage. Cases of this kind are treated in Chapter
VII below.

Nonserial multistage systems important in the chemical industries represent

another generalization, characterized by branches or loops in the flow diagram.
A third extension of the N-stage serial multistage structure is an infinite-stage
system where N tends to infinity. This case, relevant to some applications, is treated
below in Chapter VIII.
The advantage of dynamic programming as a procedure for solving optimization
problems is the simplification obtained by decomposition. It is often simpler and easier to
solve a series of single-variable stage optimization problems, and in some cases this is the
only possible procedure because simultaneous solution is mathematically or
computationally difficult or downright impossible. Certain classes of optimization
problems, however, such as linear programming problems are more efficiently solved by
special algorithms without decomposition, e.g. the simplex method.
Dynamic programming has been particularly successful in its discrete version.
Tabular computations are well suited for computer solution and can be used to handle
problems involving irregular functions, and maybe the only practicable way of solving a
problem in which the variables are required to be integers. As a rough illustration of the
computational advantages of DP, consider an optimization problem in N decision
variables, each of which can assume m alternative discrete values (e.g. 0, 1, 2, , m 1).
If, for lack of other methods of solution, we had to solve the problem by total
enumeration, we would have to examine each of mN alternative solutions for feasibility
and optimality, whereas in a dynamic programming procedure assuming that
decomposition is possible the number of alternatives to be enumerated would be
reduced to mN, namely m for each of the N stages. Thus, roughly speaking, the
computational labour increases exponentially with the number of decision variables in the
case of total exponentially with the number of decision variables in the case of total
enumeration, but only proportionately if the problem is decomposed. For large values of
m and N the computational advantages become enormous; for example, in a problem with
20 variables each of which can assume integral values from 0 to 9 we have mN = 1020

an astronomic number as against mN = 200 possible solutions if the problem is

reformulated as a 20-stage DP problem.
APPLICATIONS OF DYNAMIC PROGRAMMING
The Shortest Path through a Network
Perhaps the simplest and most straightforward application of dynamic
programming is the determination of the shortest path or whole through a network.
Consider the (stylized) road map shown in Fig. 144. A driver wants to find the
shortest route from point P to point Q. There are six intermediate junctions A, B, , F.
The lengths of all existing road sections connecting two points in the area are indicated
on the map. Any unbroken chain of road sections starting at P and ending at Q represents
a possible route through this network of roads.
A
10

C
5

Q
6

12
D

4
4

Figure 14
Assuming that the direction of travel is always from left to right having arrived at,
say, B the driver never travels back to A or D but proceeds to either C or F the number
of possible routes is finite. The problem can therefore be solved by enumerating the
alternative routes and comparing their total lengths.
Any route from P to Q is the result of three successive decisions. Starting at point
P, the driver must decide whether he will go to A or to D. If he chooses to drive to A, say,
he can proceed from A to either B or E, each of which in turn leaves him with two
alternatives, C and F. Having arrived at either C or F, he has no choice but no proceed to
the destination, Q. Since each decision is a choice between two alternatives and there are
three consecutive decisions to be made, there are 23 = 8 possible combinations, i.e. 8
alternative routes. The alternative decisions and the resulting routes can be illustrated
graphically by a decision tree as shown in Fig. 15, where the root represents the starting

point P and the branches are composed of road sections (with lengths indicated).
Comparing the total lengths from root to top, it will be seen that PDBFQ is the shortest
route, the total length being 12 + 2 + 3+ 4 = 21.

Figure 15

x1 = A or D
y0 = p

y1 = x1

y2 = x2

r1 (y0, x1)

x4 = Q

x3 = C or F

x2 = B or E

y3 = x3

r2 (y1, x2)

y4 = Q

r3 (y2, x3)

r4 (y3, x4)

Figure 16
This decision structure clearly represents a multistage decision system. The
drivers geographical position the points on the map represents the state of the
system, which is to be changed from the given initial state P to the given final state Q
through a sequence of stage decisions in such a way as to minimize the total return, i.e.,
the total distance covered. Starting from point P, i.e., the initial state y 0 = P, the two
alternatives open to the driver can be represented by two values of a decision variable x1 :
x1 = A (i.e., drive to point A) and x1 = D (drive to D). If he chooses x1 = A, he will get to
this point so that the output state of the first stage will be y 1 = A and the corresponding
return r1 (P, A) will be the distance PA = 10. Similarly, x1 D leads to y1 = D and the
return will be r1 (P, D) = 12. Proceeding in this fashion, the problem can be represented
by a four-stage decision structure as shown in Fig. 16. There is no choice at stage 4 since
the destination i.e., the final state y4 = Q = is given.
The transformation functions are yn = xn (n = 1, 2, 3, 4). The return functions rn =
rn (yn-1, xn) can be written in tabular form as follows
r1 (y0, x1)

r2 (y1, x2)

r3 (y2, x3)

r4 (y3, x4)

y0 \ x1 A

y1 \ x2 B

y2 \ x3 C

y3 \ x4 Q

12

A
D

7
2

B
E

3
4

C
F

10

6
2

3
6

5
4

Using backward recursion, the problem is solved by tabular computations as

follows:

Stage 4
y3 \ x4

f1 = r4 (y3, r4)
Q

F1 (y3)

x4 (y3)

y4 (y3)

C
F

5 (= CQ)
4 (= FQ)

5
4

Q
Q

Q
Q

Stage 3
y2 \ x3

f2 = r3 (y2, x3) + F1 (y3)

C
F

F2 (y2)

x3 (y2)

y3 (y2)

B
E

3+5
6+5

7
8

F
F

F
F

Stage 2
y1 \ x2

f3 = r2 (y1, x2) + F2 (y2)

B
E

F3 (y1)

x2 (y1)

y2 (y1)

A
D

6+7
2+7

13
9

B
B

B
B

Stage 1
y0 \ x1

f4 = r1 (y0, x1) + F3 (y1)

A
D

F4 (y0)

x1 (y0)

y1 (y0)

10 + 13

21

3+4
4+4

7+8
2+8

12 + 9

Starting with y0 = P, the solution is determined by the recursive equation system

formed by the parametric stage solutions xn = xn (yn 1) and the transformations yn = xn.
The last table gives x1 (y0) = D and y1 = x1 = D; for y1 = D, the table for stage 2 gives x2 =
B, y2 = B; and so on. The optimal values italicized in the tables are
n

yn

xn

The optimal sequence of states y0, y1, , y4 indicates that the shortest path
through the network is PDBFQ; the total length of this route is F4 (y0) = 21.
This solution procedure can be translated into a graphical method, making use of a
decision tree like that of Fig. 15. After making the three first decisions the driver finds
himself at point C or F. No matter how he got there he will have to proceed to Q, so we
draw the eight top branches of the tree; none of them can be eliminated at this stage since
the preceding decisions have not yet been determined.
Going one stage back, the first two decisions have taken the driver to either B or
E. If he has arrived at B, he can get to his destination Q either through C or F; the best
course is to go to F since BFQ = 3 + 4 = 7 whereas BCQ = 3 + 5 = 8. Therefore, no
matter which way he may have got to B, he will never proceed to C, so the branches from
B to C can be eliminated and we need only draw the branches going from B to F. This is
an application of Bellmans principle of optimality: if the optimal route from P to Q
passes through B, the remaining part of the route (from B to Q) must also be optimal; the
optimal route from P to Q cannot contain BCQ because there is a shorter route from B to
the destination. Similarly, if the driver is at point E after the first two decisions, he will
proceed to F because EFQ = 4 + 4 + 8 < ECQ = 6 + 5 = 11, so ECQ can be eliminated.
Applying a similar reasoning to the state attained after the first decision, it is seen
that the branches AE and DE need not be drawn; the shortest path from A to Q is ABFQ
= 6 + 7 = 13 and the shortest path from D to Q is DBFQ = 2 + 7 = 9. Finally, at the
starting point the choice between PA + ABFQ = 10 + 13 = 23 and PD + DBFQ = 12 + 9
= 21; the latter alternative represents the shortest total route and there is no need to draw
the branch PA.